Search Results

Search found 6387 results on 256 pages for 'cpu allocation'.

Page 94/256 | < Previous Page | 90 91 92 93 94 95 96 97 98 99 100 101  | Next Page >

  • Option Trading: Getting the most out of the event session options

    - by extended_events
    You can control different aspects of how an event session behaves by setting the event session options as part of the CREATE EVENT SESSION DDL. The default settings for the event session options are designed to handle most of the common event collection situations so I generally recommend that you just use the defaults. Like everything in the real world though, there are going to be a handful of “special cases” that require something different. This post focuses on identifying the special cases and the correct use of the options to accommodate those cases. There is a reason it’s called Default The default session options specify a total event buffer size of 4 MB with a 30 second latency. Translating this into human terms; this means that our default behavior is that the system will start processing events from the event buffer when we reach about 1.3 MB of events or after 30 seconds, which ever comes first. Aside: What’s up with the 1.3 MB, I thought you said the buffer was 4 MB?The Extended Events engine takes the total buffer size specified by MAX_MEMORY (4MB by default) and divides it into 3 equally sized buffers. This is done so that a session can be publishing events to one buffer while other buffers are being processed. There are always at least three buffers; how to get more than three is covered later. Using this configuration, the Extended Events engine can “keep up” with most event sessions on standard workloads. Why is this? The fact is that most events are small, really small; on the order of a couple hundred bytes. Even when you start considering events that carry dynamically sized data (eg. binary, text, etc.) or adding actions that collect additional data, the total size of the event is still likely to be pretty small. This means that each buffer can likely hold thousands of events before it has to be processed. When the event buffers are finally processed there is an economy of scale achieved since most targets support bulk processing of the events so they are processed at the buffer level rather than the individual event level. When all this is working together it’s more likely that a full buffer will be processed and put back into the ready queue before the remaining buffers (remember, there are at least three) are full. I know what you’re going to say: “My server is exceptional! My workload is so massive it defies categorization!” OK, maybe you weren’t going to say that exactly, but you were probably thinking it. The point is that there are situations that won’t be covered by the Default, but that’s a good place to start and this post assumes you’ve started there so that you have something to look at in order to determine if you do have a special case that needs different settings. So let’s get to the special cases… What event just fired?! How about now?! Now?! If you believe the commercial adage from Heinz Ketchup (Heinz Slow Good Ketchup ad on You Tube), some things are worth the wait. This is not a belief held by most DBAs, particularly DBAs who are looking for an answer to a troubleshooting question fast. If you’re one of these anxious DBAs, or maybe just a Program Manager doing a demo, then 30 seconds might be longer than you’re comfortable waiting. If you find yourself in this situation then consider changing the MAX_DISPATCH_LATENCY option for your event session. This option will force the event buffers to be processed based on your time schedule. This option only makes sense for the asynchronous targets since those are the ones where we allow events to build up in the event buffer – if you’re using one of the synchronous targets this option isn’t relevant. Avoid forgotten events by increasing your memory Have you ever had one of those days where you keep forgetting things? That can happen in Extended Events too; we call it dropped events. In order to optimizes for server performance and help ensure that the Extended Events doesn’t block the server if to drop events that can’t be published to a buffer because the buffer is full. You can determine if events are being dropped from a session by querying the dm_xe_sessions DMV and looking at the dropped_event_count field. Aside: Should you care if you’re dropping events?Maybe not – think about why you’re collecting data in the first place and whether you’re really going to miss a few dropped events. For example, if you’re collecting query duration stats over thousands of executions of a query it won’t make a huge difference to miss a couple executions. Use your best judgment. If you find that your session is dropping events it means that the event buffer is not large enough to handle the volume of events that are being published. There are two ways to address this problem. First, you could collect fewer events – examine you session to see if you are over collecting. Do you need all the actions you’ve specified? Could you apply a predicate to be more specific about when you fire the event? Assuming the session is defined correctly, the next option is to change the MAX_MEMORY option to a larger number. Picking the right event buffer size might take some trial and error, but a good place to start is with the number of dropped events compared to the number you’ve collected. Aside: There are three different behaviors for dropping events that you specify using the EVENT_RETENTION_MODE option. The default is to allow single event loss and you should stick with this setting since it is the best choice for keeping the impact on server performance low.You’ll be tempted to use the setting to not lose any events (NO_EVENT_LOSS) – resist this urge since it can result in blocking on the server. If you’re worried that you’re losing events you should be increasing your event buffer memory as described in this section. Some events are too big to fail A less common reason for dropping an event is when an event is so large that it can’t fit into the event buffer. Even though most events are going to be small, you might find a condition that occasionally generates a very large event. You can determine if your session is dropping large events by looking at the dm_xe_sessions DMV once again, this time check the largest_event_dropped_size. If this value is larger than the size of your event buffer [remember, the size of your event buffer, by default, is max_memory / 3] then you need a large event buffer. To specify a large event buffer you set the MAX_EVENT_SIZE option to a value large enough to fit the largest event dropped based on data from the DMV. When you set this option the Extended Events engine will create two buffers of this size to accommodate these large events. As an added bonus (no extra charge) the large event buffer will also be used to store normal events in the cases where the normal event buffers are all full and waiting to be processed. (Note: This is just a side-effect, not the intended use. If you’re dropping many normal events then you should increase your normal event buffer size.) Partitioning: moving your events to a sub-division Earlier I alluded to the fact that you can configure your event session to use more than the standard three event buffers – this is called partitioning and is controlled by the MEMORY_PARTITION_MODE option. The result of setting this option is fairly easy to explain, but knowing when to use it is a bit more art than science. First the science… You can configure partitioning in three ways: None, Per NUMA Node & Per CPU. This specifies the location where sets of event buffers are created with fairly obvious implication. There are rules we follow for sub-dividing the total memory (specified by MAX_MEMORY) between all the event buffers that are specific to the mode used: None: 3 buffers (fixed)Node: 3 * number_of_nodesCPU: 2.5 * number_of_cpus Here are some examples of what this means for different Node/CPU counts: Configuration None Node CPU 2 CPUs, 1 Node 3 buffers 3 buffers 5 buffers 6 CPUs, 2 Node 3 buffers 6 buffers 15 buffers 40 CPUs, 5 Nodes 3 buffers 15 buffers 100 buffers   Aside: Buffer size on multi-processor computersAs the number of Nodes or CPUs increases, the size of the event buffer gets smaller because the total memory is sub-divided into more pieces. The defaults will hold up to this for a while since each buffer set is holding events only from the Node or CPU that it is associated with, but at some point the buffers will get too small and you’ll either see events being dropped or you’ll get an error when you create your session because you’re below the minimum buffer size. Increase the MAX_MEMORY setting to an appropriate number for the configuration. The most likely reason to start partitioning is going to be related to performance. If you notice that running an event session is impacting the performance of your server beyond a reasonably expected level [Yes, there is a reasonably expected level of work required to collect events.] then partitioning might be an answer. Before you partition you might want to check a few other things: Is your event retention set to NO_EVENT_LOSS and causing blocking? (I told you not to do this.) Consider changing your event loss mode or increasing memory. Are you over collecting and causing more work than necessary? Consider adding predicates to events or removing unnecessary events and actions from your session. Are you writing the file target to the same slow disk that you use for TempDB and your other high activity databases? <kidding> <not really> It’s always worth considering the end to end picture – if you’re writing events to a file you can be impacted by I/O, network; all the usual stuff. Assuming you’ve ruled out the obvious (and not so obvious) issues, there are performance conditions that will be addressed by partitioning. For example, it’s possible to have a successful event session (eg. no dropped events) but still see a performance impact because you have many CPUs all attempting to write to the same free buffer and having to wait in line to finish their work. This is a case where partitioning would relieve the contention between the different CPUs and likely reduce the performance impact cause by the event session. There is no DMV you can check to find these conditions – sorry – that’s where the art comes in. This is  largely a matter of experimentation. On the bright side you probably won’t need to to worry about this level of detail all that often. The performance impact of Extended Events is significantly lower than what you may be used to with SQL Trace. You will likely only care about the impact if you are trying to set up a long running event session that will be part of your everyday workload – sessions used for short term troubleshooting will likely fall into the “reasonably expected impact” category. Hey buddy – I think you forgot something OK, there are two options I didn’t cover: STARTUP_STATE & TRACK_CAUSALITY. If you want your event sessions to start automatically when the server starts, set the STARTUP_STATE option to ON. (Now there is only one option I didn’t cover.) I’m going to leave causality for another post since it’s not really related to session behavior, it’s more about event analysis. - Mike Share this post: email it! | bookmark it! | digg it! | reddit! | kick it! | live it!

    Read the article

  • Is recursion really bad?

    - by dotneteer
    After my previous post about the stack space, it appears that there is perception from the feedback that recursion is bad and we should avoid deep recursion. After writing a compiler, I know that the modern computer and compiler are complex enough and one cannot automatically assume that a hand crafted code would out-perform the compiler optimization. The only way is to do some prototype to find out. So why recursive code may not perform as well? Compilers place frames on a stack. In additional to arguments and local variables, compiles also need to place frame and program pointers on the frame, resulting in overheads. So why hand-crafted code may not performance as well? The stack used by a compiler is a simpler data structure and can grow and shrink cleanly. To replace recursion with out own stack, our stack is allocated in the heap that is far more complicated to manage. There could be overhead as well if the compiler needs to mark objects for garbage collection. Compiler also needs to worry about the memory fragmentation. Then there is additional complexity: CPUs have registers and multiple levels of cache. Register access is a few times faster than in-CPU cache access and is a few 10s times than on-board memory access. So it is up to the OS and compiler to maximize the use of register and in-CPU cache. For my particular problem, I did an experiment to rewrite my c# version of recursive code with a loop and stack approach. So here are the outcomes of the two approaches:   Recursive call Loop and Stack Lines of code for the algorithm 17 46 Speed Baseline 3% faster Readability Clean Far more complex So at the end, I was able to achieve 3% better performance with other drawbacks. My message is never assuming your sophisticated approach would automatically work out better than a simpler approach with a modern computer and compiler. Gage carefully before committing to a more complex approach.

    Read the article

  • system overrides my mount parameters in /etc/fstab

    - by valya
    [.../~]$ mount /dev/sda4 on / type ext4 (rw,commit=60,commit=0) [.../~]$ cat /etc/fstab # UNCONFIGURED FSTAB FOR BASE SYSTEM UUID=70739c04-fcb6-4747-803c-824f9c894f41 / ext4 defaults,commit=60 0 1 What can I do about it? It seems strange. I want to be able to set any commit time I want Edit: added /proc/mounts contents [.../~]$ cat /proc/mounts rootfs / rootfs rw 0 0 none /sys sysfs rw,nosuid,nodev,noexec,relatime 0 0 none /proc proc rw,nosuid,nodev,noexec,relatime 0 0 none /dev devtmpfs rw,relatime,size=886332k,nr_inodes=221583,mode=755 0 0 none /dev/pts devpts rw,nosuid,noexec,relatime,gid=5,mode=620,ptmxmode=000 0 0 fusectl /sys/fs/fuse/connections fusectl rw,relatime 0 0 /dev/disk/by-uuid/70739c04-fcb6-4747-803c-824f9c894f41 / ext4 rw,relatime,barrier=1,data=ordered 0 0 none /sys/kernel/debug debugfs rw,relatime 0 0 none /sys/kernel/security securityfs rw,relatime 0 0 none /dev/shm tmpfs rw,nosuid,nodev,relatime 0 0 none /var/run tmpfs rw,nosuid,relatime,mode=755 0 0 none /var/lock tmpfs rw,nosuid,nodev,noexec,relatime 0 0 /dev/sda3 /media/megahard fuseblk rw,nosuid,nodev,relatime,user_id=0,group_id=0,allow_other,blksize=4096 0 0 cgroup /dev/cgroup/cpu cgroup rw,relatime,cpu,release_agent=/usr/local/sbin/cgroup_clean 0 0 gvfs-fuse-daemon /home/va1en0k/.gvfs fuse.gvfs-fuse-daemon rw,nosuid,nodev,relatime,user_id=1000,group_id=1000 0 0

    Read the article

  • Option Trading: Getting the most out of the event session options

    - by extended_events
    You can control different aspects of how an event session behaves by setting the event session options as part of the CREATE EVENT SESSION DDL. The default settings for the event session options are designed to handle most of the common event collection situations so I generally recommend that you just use the defaults. Like everything in the real world though, there are going to be a handful of “special cases” that require something different. This post focuses on identifying the special cases and the correct use of the options to accommodate those cases. There is a reason it’s called Default The default session options specify a total event buffer size of 4 MB with a 30 second latency. Translating this into human terms; this means that our default behavior is that the system will start processing events from the event buffer when we reach about 1.3 MB of events or after 30 seconds, which ever comes first. Aside: What’s up with the 1.3 MB, I thought you said the buffer was 4 MB?The Extended Events engine takes the total buffer size specified by MAX_MEMORY (4MB by default) and divides it into 3 equally sized buffers. This is done so that a session can be publishing events to one buffer while other buffers are being processed. There are always at least three buffers; how to get more than three is covered later. Using this configuration, the Extended Events engine can “keep up” with most event sessions on standard workloads. Why is this? The fact is that most events are small, really small; on the order of a couple hundred bytes. Even when you start considering events that carry dynamically sized data (eg. binary, text, etc.) or adding actions that collect additional data, the total size of the event is still likely to be pretty small. This means that each buffer can likely hold thousands of events before it has to be processed. When the event buffers are finally processed there is an economy of scale achieved since most targets support bulk processing of the events so they are processed at the buffer level rather than the individual event level. When all this is working together it’s more likely that a full buffer will be processed and put back into the ready queue before the remaining buffers (remember, there are at least three) are full. I know what you’re going to say: “My server is exceptional! My workload is so massive it defies categorization!” OK, maybe you weren’t going to say that exactly, but you were probably thinking it. The point is that there are situations that won’t be covered by the Default, but that’s a good place to start and this post assumes you’ve started there so that you have something to look at in order to determine if you do have a special case that needs different settings. So let’s get to the special cases… What event just fired?! How about now?! Now?! If you believe the commercial adage from Heinz Ketchup (Heinz Slow Good Ketchup ad on You Tube), some things are worth the wait. This is not a belief held by most DBAs, particularly DBAs who are looking for an answer to a troubleshooting question fast. If you’re one of these anxious DBAs, or maybe just a Program Manager doing a demo, then 30 seconds might be longer than you’re comfortable waiting. If you find yourself in this situation then consider changing the MAX_DISPATCH_LATENCY option for your event session. This option will force the event buffers to be processed based on your time schedule. This option only makes sense for the asynchronous targets since those are the ones where we allow events to build up in the event buffer – if you’re using one of the synchronous targets this option isn’t relevant. Avoid forgotten events by increasing your memory Have you ever had one of those days where you keep forgetting things? That can happen in Extended Events too; we call it dropped events. In order to optimizes for server performance and help ensure that the Extended Events doesn’t block the server if to drop events that can’t be published to a buffer because the buffer is full. You can determine if events are being dropped from a session by querying the dm_xe_sessions DMV and looking at the dropped_event_count field. Aside: Should you care if you’re dropping events?Maybe not – think about why you’re collecting data in the first place and whether you’re really going to miss a few dropped events. For example, if you’re collecting query duration stats over thousands of executions of a query it won’t make a huge difference to miss a couple executions. Use your best judgment. If you find that your session is dropping events it means that the event buffer is not large enough to handle the volume of events that are being published. There are two ways to address this problem. First, you could collect fewer events – examine you session to see if you are over collecting. Do you need all the actions you’ve specified? Could you apply a predicate to be more specific about when you fire the event? Assuming the session is defined correctly, the next option is to change the MAX_MEMORY option to a larger number. Picking the right event buffer size might take some trial and error, but a good place to start is with the number of dropped events compared to the number you’ve collected. Aside: There are three different behaviors for dropping events that you specify using the EVENT_RETENTION_MODE option. The default is to allow single event loss and you should stick with this setting since it is the best choice for keeping the impact on server performance low.You’ll be tempted to use the setting to not lose any events (NO_EVENT_LOSS) – resist this urge since it can result in blocking on the server. If you’re worried that you’re losing events you should be increasing your event buffer memory as described in this section. Some events are too big to fail A less common reason for dropping an event is when an event is so large that it can’t fit into the event buffer. Even though most events are going to be small, you might find a condition that occasionally generates a very large event. You can determine if your session is dropping large events by looking at the dm_xe_sessions DMV once again, this time check the largest_event_dropped_size. If this value is larger than the size of your event buffer [remember, the size of your event buffer, by default, is max_memory / 3] then you need a large event buffer. To specify a large event buffer you set the MAX_EVENT_SIZE option to a value large enough to fit the largest event dropped based on data from the DMV. When you set this option the Extended Events engine will create two buffers of this size to accommodate these large events. As an added bonus (no extra charge) the large event buffer will also be used to store normal events in the cases where the normal event buffers are all full and waiting to be processed. (Note: This is just a side-effect, not the intended use. If you’re dropping many normal events then you should increase your normal event buffer size.) Partitioning: moving your events to a sub-division Earlier I alluded to the fact that you can configure your event session to use more than the standard three event buffers – this is called partitioning and is controlled by the MEMORY_PARTITION_MODE option. The result of setting this option is fairly easy to explain, but knowing when to use it is a bit more art than science. First the science… You can configure partitioning in three ways: None, Per NUMA Node & Per CPU. This specifies the location where sets of event buffers are created with fairly obvious implication. There are rules we follow for sub-dividing the total memory (specified by MAX_MEMORY) between all the event buffers that are specific to the mode used: None: 3 buffers (fixed)Node: 3 * number_of_nodesCPU: 2.5 * number_of_cpus Here are some examples of what this means for different Node/CPU counts: Configuration None Node CPU 2 CPUs, 1 Node 3 buffers 3 buffers 5 buffers 6 CPUs, 2 Node 3 buffers 6 buffers 15 buffers 40 CPUs, 5 Nodes 3 buffers 15 buffers 100 buffers   Aside: Buffer size on multi-processor computersAs the number of Nodes or CPUs increases, the size of the event buffer gets smaller because the total memory is sub-divided into more pieces. The defaults will hold up to this for a while since each buffer set is holding events only from the Node or CPU that it is associated with, but at some point the buffers will get too small and you’ll either see events being dropped or you’ll get an error when you create your session because you’re below the minimum buffer size. Increase the MAX_MEMORY setting to an appropriate number for the configuration. The most likely reason to start partitioning is going to be related to performance. If you notice that running an event session is impacting the performance of your server beyond a reasonably expected level [Yes, there is a reasonably expected level of work required to collect events.] then partitioning might be an answer. Before you partition you might want to check a few other things: Is your event retention set to NO_EVENT_LOSS and causing blocking? (I told you not to do this.) Consider changing your event loss mode or increasing memory. Are you over collecting and causing more work than necessary? Consider adding predicates to events or removing unnecessary events and actions from your session. Are you writing the file target to the same slow disk that you use for TempDB and your other high activity databases? <kidding> <not really> It’s always worth considering the end to end picture – if you’re writing events to a file you can be impacted by I/O, network; all the usual stuff. Assuming you’ve ruled out the obvious (and not so obvious) issues, there are performance conditions that will be addressed by partitioning. For example, it’s possible to have a successful event session (eg. no dropped events) but still see a performance impact because you have many CPUs all attempting to write to the same free buffer and having to wait in line to finish their work. This is a case where partitioning would relieve the contention between the different CPUs and likely reduce the performance impact cause by the event session. There is no DMV you can check to find these conditions – sorry – that’s where the art comes in. This is  largely a matter of experimentation. On the bright side you probably won’t need to to worry about this level of detail all that often. The performance impact of Extended Events is significantly lower than what you may be used to with SQL Trace. You will likely only care about the impact if you are trying to set up a long running event session that will be part of your everyday workload – sessions used for short term troubleshooting will likely fall into the “reasonably expected impact” category. Hey buddy – I think you forgot something OK, there are two options I didn’t cover: STARTUP_STATE & TRACK_CAUSALITY. If you want your event sessions to start automatically when the server starts, set the STARTUP_STATE option to ON. (Now there is only one option I didn’t cover.) I’m going to leave causality for another post since it’s not really related to session behavior, it’s more about event analysis. - Mike Share this post: email it! | bookmark it! | digg it! | reddit! | kick it! | live it!

    Read the article

  • XP Mode (Windows Virtual PC for Windows 7) no longer requires hardware virtualisation - hurrah !

    - by Liam Westley
    Windows Virtual PC (aka XP Mode) When XP Mode was released, it insisted on hardware virtualisation being present on your CPU and enabled in the BIOS.  Given that Windows Virtual PC was based on an improved Virtual PC 2007, which provided hardware virtualisation as a user selectable option, I did wonder why on earth Microsoft thought this was a good idea.  Not only do many people not have a CPU with hardware virtualisation support, some manufacturers don't provide a BIOS option to enable this setting, especially on laptops - yes Sony, Toshiba and Acer, I'm looking at you. Dumb and dumber This issue became a double whammy; not only was Microsoft a bit dumb on not supporting Windows Virtual PC without hardware virtualisation, your hardware manufacturer was also dumb in not supporting the option in the BIOS. Microsoft update to Windows Virtual PC Belatedly, Microsoft has seen the problem with this hardware virtualisation requirement and has now released a new version of Windows Virtual PC that works without hardware virtualisation.  This is really good news for those with older (or limited) CPUs and rubbish BIOS firmware. You can details of how to download the new versions of XP Mode here, http://blogs.msdn.com/virtual_pc_guy/archive/2010/03/18/windows-virtual-pc-no-hardware-virtualization-update-now-available-for-download.aspx And there is also an explanation of why the hardware virtualisation requirement was in place for previous releases, http://blogs.msdn.com/virtual_pc_guy/archive/2010/03/18/windows-virtual-pc-now-without-the-need-for-hardware-virtualization.aspx

    Read the article

  • Why can I not map a dynamic texture in D3D?

    - by sebf
    I am trying to map a Texture2D resource in DirectX11 via SharpDX. The resource is declared as a ShaderResource, with Dynamic usage and the 'Write' CPU flag specified. My call however fails with a generic exception from SharpDX: _Parent.Context.MapSubresource( _Resource, 0, SharpDX.Direct3D11.MapMode.Write, SharpDX.Direct3D11.MapFlags.None, out stream ); I see from this question that it is supported. The MSDN docs and this other question hint that instead of using Context.MapSubresource() I should be using Texture2D.Map(), however, the DirectX11 Texture2D class does not define Map() (though it does for the D3D 10 equivalent). If I call the above with MapMode.WriteDiscard, the call succeeds but in this case the previous content of the texture is lost, which is no good when I only want to update a section of it. Has the Map() method been removed in Direct3D 11 or am I looking in the wrong place? Is the MapSubresource() method unsuitable or am I using it wrong? EDIT: I declared my resource as Dynamic with CPU Write Flags - not Default as I originaly wrote - sorry for the fairly huge 'typo' that changes the entire question!

    Read the article

  • How to stop Cairo Dock minimizing Conky on Show Desktop?

    - by César
    Every time I use Cairo Dock Show Desktop add-on Conky minimizes: I've read about the own_window_type override option on .conkyrc and it seems to work for some people but it doesn't work for me. Conky won't show up if I use this option (it is currently set to own_window_type normal). Any suggestions? .conkyrc # Conky settings # background no update_interval 1 cpu_avg_samples 2 net_avg_samples 2 override_utf8_locale yes double_buffer yes no_buffers yes text_buffer_size 2048 #imlib_cache_size 0 temperature_unit fahrenheit # Window specifications # own_window yes own_window_type normal own_window_transparent yes own_window_hints undecorate,sticky,skip_taskbar,skip_pager,below border_inner_margin 0 border_outer_margin 0 minimum_size 200 250 maximum_width 200 alignment tr gap_x 35 gap_y 55 # Graphics settings # draw_shades no draw_outline no draw_borders no draw_graph_borders no # Text settings # use_xft yes override_utf8_locale yes xftfont Neuropolitical:size=8 xftalpha 0.8 uppercase no temperature_unit celsius default_color FFFFFF # Lua Load # lua_load ~/.lua/scripts/clock_rings.lua lua_draw_hook_pre clock_rings TEXT ${font Neuropolitical:size=42}${time %e} ${goto 100}${font Neuropolitical:size=18}${color FF3300}${voffset -75}${time %b} ${font Neuropolitical:size=10}${color FF3300}${voffset 15}${time %A}${color FF3300}${hr} ${goto 100}${font Neuropolitical:size=15}${color FFFFFF}${voffset -35}${time %Y} ${font Neuropolitical:size=30}${voffset 40}${alignc}${time %H}:${time %M} ${goto 175}${voffset -30}${font Neuropolitical:size=10}${time %S} ${voffset 10}${font Neuropolitical:size=11}${color FF3300}${alignr}HOME${font} ${font Neuropolitical:size=13}${color FFFFFF}${alignr}temp: ${weather http://weather.noaa.gov/pub/data/observations/metar/stations/ LQBK temperature temperature 30} °C${font} ${hr} ${image ~/.conky/logo.png -p 165,10 -s 35x35} ${color FFFFFF}${font Neuropolitical:size=8}Uptime: ${uptime_short} ${color FFFFFF}${font Neuropolitical:size=8}Processes: ${processes} ${color FFFFFF}${font Neuropolitical:size=8}Running: ${running_processes} ${color FF3300}${goto 125}${voffset 27}CPU ${color FFFFFF}${goto 125}${cpu cpu0}% ${color FF3300}${goto 125}${voffset 55}RAM ${color FFFFFF}${goto 125}${memperc}% ${color FF3300}${goto 125}${voffset 56}Swap ${color FFFFFF}${goto 125}${swapperc}% ${color FF3300}${goto 125}${voffset 57}Disk ${color FFFFFF}${goto 125}${fs_used_perc /}% ${color FF3300}${goto 130}${voffset 55}Net ${color FFFFFF}${goto 130}${downspeed eth0} ${color FFFFFF}${goto 130}${upspeed eth0} ${color FF3300}${font Neuropolitical:size=8}${alignr}${nodename} ${color FF3300}${font Neuropolitical:size=8}${alignr}${pre_exec cat /etc/issue.net} $machine ${color FF3300}${font Neuropolitical:size=8}${alignr}Kernel: ${kernel} ${hr}

    Read the article

  • Whole Lotta Virtualization Goin' On

    - by rickramsey
    Lately we've published a lot of content about virtualization. Here's a sampling. Podcat: Technology Preview of Transcendent Memory Turns out that in a virtual environment, RAM is the bottleneck. Not because it's slow, it's not, but because each CPU still had to use its own RAM. Which gets expensive. In this podcast, Dan Magenheimer describes how Oracle and the open source community taught the guest kernel in Oracle Linux to share its memory with other CPU's. Transcendent memory will wind up saving large data centers a lot of money. Find out how. Tech Article: How to Use Oracle VM Templates This article describes how to prepare an Oracle VM environment to use Oracle VM Templates, how to obtain a template, and how to deploy the template to your Oracle VM environment. It also describes how to create a virtual machine based on that template and how you can clone the template and change the clone's configuration. Tech Article: How to Set Up a Load Balanced Application Across Two Oracle Solaris Zones Install Apache Tomcat on two Oracle Solaris zones. Connect them across a VPN. And let the Integrated Load Balancer in Oracle Solaris 11 manage traffic. Presto: high(er) availability in a single server. Tech Article: How to Install Oracle RAC on Oracle Solaris Zone Clusters Learn how to implement a multi-tiered database environment that isolates database tiers and administrative domains, while taking advantage of centralized (and simpler) cluster admin. For fans of Jerry Lee Lewis If you're a fan of Jerry Lee Lewis, you might enjoy this video. - Rick Website Newsletter Facebook Twitter

    Read the article

  • Apache VERY high page load time

    - by Aaron Waller
    My Drupal 6 site has been running smoothly for years but recently has experienced intermittent periods of extreme slowness (10-60 sec page loads). Several hours of slowness followed by hours of normal (4-6 sec) page loads. The page always loads with no error, just sometimes takes forever. My setup: Windows Server 2003 Apache/2.2.15 (Win32) Jrun/4.0 PHP 5 MySql 5.1 Drupal 6 Cold fusion 9 Vmware virtual environment DMZ behind a corporate firewall Traffic: 1-3 hits/sec avg Troubleshooting No applicable errors in apache error log No errors in drupal event log Drupal devel module shows 242 queries in 366.23 milliseconds,page execution time 2069.62 ms. (So it looks like queries and php scripts are not the problem) NO unusually high CPU, memory, or disk IO Cold fusion apps, and other static pages outside of drupal also load slow webpagetest.org test shows very high time-to-first-byte The problem seems to be with Apache responding to requests, but previously I've only seen this behavior under 100% cpu load. Judging solely by resource monitoring, it looks as though very little is going on. Here is the kicker - roughly half of the site's access comes from our LAN, but if I disable the firewall rule and block access from outside of our network, internal (LAN) access (1000+ devices) is speedy. But as soon as outside access is restored the site is crippled. Apache config? Crawlers/bots? Attackers? I'm at the end of my rope, where should I be looking to determine where the problem lies?

    Read the article

  • How do I measure performance of a virtual server?

    - by Sergey
    I've got a VPS running Ubuntu. Being a virtual server, I understand that it shares resources with unknown number of other servers, and I'm noticing that it's considerably slower than my desktop machine. Is there some tool to measure the performance of the virtual machine? I'd be curious to see some approximate measure similar to bogomips, possibly for CPU (operations/sec), memory and disk read/write speed. I'd like to be able to compare those numbers to my desktop machine. I'm not interested in the specs of the actual physical machine my VPS is running on - by doing cat /proc/cpuinfo I can see that it's a nice quad-core Xeon machine, but it doesn't matter to me. I'm basically interested in how fast a program would run in my VPS - how many CPU operations it can make in a second, how many bytes to write to RAM or to disk. I only have ssh access to the machine so the tool need to be command-line. I could write a script which, say, does some calculations in a loop for a second and counts how many loops it was able to do, or something similar to measure disk and RAM performance. But I'm sure something like this already exists.

    Read the article

  • Instructor Insight: Using the Container Database in Oracle Database 12 c

    - by Breanne Cooley
    The first time I examined the Oracle Database 12c architecture, I wasn’t quite sure what I thought about the Container Database (CDB). In the current release of the Oracle RDBMS, the administrator now has a choice of whether or not to employ a CDB. Bundling Databases Inside One Container In today’s IT industry, consolidation is a common challenge. With potentially hundreds of databases to manage and maintain, an administrator will require a great deal of time and resources to upgrade and patch software. Why not consider deploying a container database to streamline this activity? By “bundling” several databases together inside one container, in the form of a pluggable database, we can save on overhead process resources and CPU time. Furthermore, we can reduce the human effort required for periodically patching and maintaining the software. Minimizing Storage Most IT professionals understand the concept of storage, as in solid state or non-rotating. Let’s take one-to-many databases and “plug” them into ONE designated container database. We can minimize many redundant pieces that would otherwise require separate storage and architecture, as was the case in previous releases of the Oracle RDBMS. The data dictionary can be housed and shared in one CDB, with individual metadata content for each pluggable database. We also won’t need as many background processes either, thus reducing the overhead cost of the CPU resource. Improve Security Levels within Each Pluggable Database  We can now segregate the CDB-administrator role from that of the pluggable-database administrator as well, achieving improved security levels within each pluggable database and within the CDB. And if the administrator chooses to use the non-CDB architecture, everything is backwards compatible, too.  The bottom line: it's a good idea to at least consider using a CDB. -Christopher Andrews, Senior Principal Instructor, Oracle University

    Read the article

  • How do I get my Lexmark x4650 printer working?

    - by Fallen Dohingy
    I think that my printer stopped working with the switch to gnome 3 or unity. Yes I have tried 32 and 64 bit os's. Here is the driver In order to actually install the driver, you need to extract it and then open up terminal and type sudo and then a space. Then drag the script into the terminal window. Here is what it said in the diver install window: Extracting file: printdriver.te Extracting file: lexmark-08z-series-driver-1.0-1.i386.deb Extracting file: launcher.c Extracting file: launcherfallendohingy@Ubuntu-Inspiron-15R:~$ sudo '/home/fallendohingy/Downloads/lexmark-08z-series-driver-1.0-1.i386.deb.sh' [sudo] password for fallendohingy: Verifying archive integrity... All good. Uncompressing nixstaller.............................................................. Collecting info for this system... Operating system: linux CPU Arch: x86_64 Warning: No installer for "x86_64" found, defaulting to x86... TRACKING IDENT = 170209 cpu speed = 2394 MHz ram size = 3762.69921875 MB hd avail = 74348 MB (gtk:17645): GdkPixbuf-WARNING **: Cannot open pixbuf loader module file '/usr/lib/i386-linux-gnu/gdk-pixbuf-2.0/2.10.0/loaders.cache': No such file or directory (gtk:17645): GdkPixbuf-WARNING **: Cannot open pixbuf loader module file '/usr/lib/i386-linux-gnu/gdk-pixbuf-2.0/2.10.0/loaders.cache': No such file or directory (gtk:17645): GdkPixbuf-WARNING **: Cannot open pixbuf loader module file '/usr/lib/i386-linux-gnu/gdk-pixbuf-2.0/2.10.0/loaders.cache': No such file or directory (gtk:17645): GdkPixbuf-WARNING **: Cannot open pixbuf loader module file '/usr/lib/i386-linux-gnu/gdk-pixbuf-2.0/2.10.0/loaders.cache': No such file or directory /usr/lib/gio/modules/libgvfsdbus.so: wrong ELF class: ELFCLASS64 Failed to load module: /usr/lib/gio/modules/libgvfsdbus.so Extracting file: lsbrowser Extracting file: lsusbdevice Using dpkg installation ============================= Execute: dpkg -i --force-architecture lexmark-08z-series-driver-1.0-1.i386.deb > /tmp/selfgz17540/pkg/files/dpkg_msgs ============================= ============================= Execute: rm lexmark-08z-series-driver-1.0-1.i386.deb ============================= ============================= Execute: /sbin/udevadm control --reload-rules ============================= Successfully installed the .deb Lexmark drivers.

    Read the article

  • How to diagnose and fix Kernel Panic Fatal Machine Check error?

    - by 0x4a6f4672
    I have got a new Samsung Series 7 laptop with dual boot setup for Windows 8 and Ubuntu 12.10. A fine machine comparable to a Macbook Pro. The Ubuntu installation was quite a hassle, but with the help of Boot Repair finally it seemed to work. Or so I thought. Windows 8 starts fine, but if I want to start Ubuntu regularly the following Machine Check Exception error occurs, quite similar to this one [Hardware Error] CPU 1: Machine Check Exception: 5 Bank 6 [Hardware Error] RIP !inexact! 33 <00007fab2074598a> [Hardware Error] TSC 95b623464c ADDR fe400 MISC 3880000086 .. [similar messages for CPU 2,3 and 0] .. [Hardware Error] Machine Check: Processor context corrupt Kernel panic - not syncing: Fatal Machine Check Rebooting in 30 seconds Kernel panic does not sound good. Then it starts to reboot, and the second boot trial often works. Is it a Kernel or driver problem? The laptop has an Intel Core i7 processor. I already deactivated Hyperthreading in the BIOS, but it does not seem to help :-( I also disabled the Execute Disable Bit (EDB) flag in the BIOS. EDB is an Intel hardware-based security feature that can help reduce system exposure to viruses and malicious code. Since I disabled it, the error did occur less frequently, but it still appears occasionally :-( It seems to be the same error as described here and here. Maybe a Samsung specific Kernel problem? A similar error also happens on a Samsung Ultrabook Series 9 (which seems to be kernel bugs 49161 and 47121). At my Samsung Series 7, it still occurs for instance during booting on battery after "Checking battery state". Perhaps anyone else has an idea? These Kernel Panic errors are reallly annoying..

    Read the article

  • SD Card Reader not working in Ubuntu 12.04

    - by tripkane
    I have a read many other posts on this issue and believe that Ubuntu 12.04 is not even recognizing my SD Card Reader as just that: Computer Model: Metabox (Australian builder of Clevo laptops) / Clevo P150EM OS: Ubuntu 12.04 (64 Bit) CPU: Intel(R) Core(TM) i7-3720QM CPU @ 2.60GHz HD: 120GB Intel 550/520MB/s SSD According to the people who built my computer, the specs of the SD Card reader in my comp are as follows: Manufacture: Realtek Semiconduct Corp. Location: PCI bus 3 Hardware ID: PCI\Ven_10EC&DEV_5289&SUBSYS_51051558 Physical device object name: \Device\NTPNP_PCI0015 Here are the relevant outputs of the following commands run from the terminal: sudo lshw *-generic UNCLAIMED description: Unassigned class product: Realtek Semiconductor Co., Ltd. vendor: Realtek Semiconductor Co., Ltd. physical id: 0 bus info: pci@0000:03:00.0 version: 01 width: 32 bits clock: 33MHz capabilities: pm msi pciexpress msix vpd bus_master cap_list configuration: latency=0 resources: memory:f6a00000-f6a0ffff sudo lspci -v -nn 03:00.0 Unassigned class [ff00]: Realtek Semiconductor Co., Ltd. Device [10ec:5289] (rev 01) Subsystem: CLEVO/KAPOK Computer Device [1558:5105] Flags: bus master, fast devsel, latency 0, IRQ 4 Memory at f6a00000 (32-bit, non-prefetchable) [size=64K] Capabilities: [40] Power Management version 3 Capabilities: [50] MSI: Enable- Count=1/1 Maskable- 64bit+ Capabilities: [70] Express Endpoint, MSI 00 Capabilities: [b0] MSI-X: Enable- Count=1 Masked- Capabilities: [d0] Vital Product Data Capabilities: [100] Advanced Error Reporting Capabilities: [140] Virtual Channel Capabilities: [160] Device Serial Number 00-00-00-00-00-00-00-00 Does the unassigned details of these outputs mean that Ubunutu desn't know that the SD Card Reader is one and what do with it? and if so how should I go about fixing it?? Cheers ;)

    Read the article

  • Turn off Windows Defender on your builds

    - by george_v_reilly
    I've spent some time this evening profiling a Python application on Windows, trying to find out why it was so much slower than on Mac or Linux. The application is an in-house build tool which reads a number of config files, then writes some output files. Using the RunSnakeRun Python profile viewer on Windows, two things immediately leapt out at me: we were running os.stat a lot and file.close was really expensive. A quick test convinced me that we were stat-ing the same files over and over. It was a combination of explicit checks and implicit code, like os.walk calling os.path.isdir. I wrote a little cache that memoizes the results, which brought the cost of the os.stats down from 1.5 seconds to 0.6. Figuring out why closing files was so expensive was harder. I was writing 77 files, totaling just over 1MB, and it was taking 3.5 seconds. It turned out that it wasn't the UTF-8 codec or newline translation. It was simply that closing those files took far longer than it should have. I decided to try a different profiler, hoping to learn more. I downloaded the Windows Performance Toolkit. I recorded a couple of traces of my application running, then I looked at them in the Windows Performance Analyzer, whereupon I saw that in each case, the CPU spike of my app was followed by a CPU spike in MsMpEng.exe. What's MsMpEng.exe? It's Microsoft's antimalware engine, at the heart of Windows Defender. I added my build tree to the list of excluded locations, and my runtime halved. The 3.5 seconds of file closing dropped to 60 milliseconds, a 98% reduction. The moral of this story is: don't let your virus checker run on your builds.

    Read the article

  • Constant game speed independent of variable FPS in OpenGL with GLUT?

    - by Nazgulled
    I've been reading Koen Witters detailed article about different game loop solutions but I'm having some problems implementing the last one with GLUT, which is the recommended one. After reading a couple of articles, tutorials and code from other people on how to achieve a constant game speed, I think that what I currently have implemented (I'll post the code below) is what Koen Witters called Game Speed dependent on Variable FPS, the second on his article. First, through my searching experience, there's a couple of people that probably have the knowledge to help out on this but don't know what GLUT is and I'm going to try and explain (feel free to correct me) the relevant functions for my problem of this OpenGL toolkit. Skip this section if you know what GLUT is and how to play with it. GLUT Toolkit: GLUT is an OpenGL toolkit and helps with common tasks in OpenGL. The glutDisplayFunc(renderScene) takes a pointer to a renderScene() function callback, which will be responsible for rendering everything. The renderScene() function will only be called once after the callback registration. The glutTimerFunc(TIMER_MILLISECONDS, processAnimationTimer, 0) takes the number of milliseconds to pass before calling the callback processAnimationTimer(). The last argument is just a value to pass to the timer callback. The processAnimationTimer() will not be called each TIMER_MILLISECONDS but just once. The glutPostRedisplay() function requests GLUT to render a new frame so we need call this every time we change something in the scene. The glutIdleFunc(renderScene) could be used to register a callback to renderScene() (this does not make glutDisplayFunc() irrelevant) but this function should be avoided because the idle callback is continuously called when events are not being received, increasing the CPU load. The glutGet(GLUT_ELAPSED_TIME) function returns the number of milliseconds since glutInit was called (or first call to glutGet(GLUT_ELAPSED_TIME)). That's the timer we have with GLUT. I know there are better alternatives for high resolution timers, but let's keep with this one for now. I think this is enough information on how GLUT renders frames so people that didn't know about it could also pitch in this question to try and help if they fell like it. Current Implementation: Now, I'm not sure I have correctly implemented the second solution proposed by Koen, Game Speed dependent on Variable FPS. The relevant code for that goes like this: #define TICKS_PER_SECOND 30 #define MOVEMENT_SPEED 2.0f const int TIMER_MILLISECONDS = 1000 / TICKS_PER_SECOND; int previousTime; int currentTime; int elapsedTime; void renderScene(void) { (...) // Setup the camera position and looking point SceneCamera.LookAt(); // Do all drawing below... (...) } void processAnimationTimer(int value) { // setups the timer to be called again glutTimerFunc(TIMER_MILLISECONDS, processAnimationTimer, 0); // Get the time when the previous frame was rendered previousTime = currentTime; // Get the current time (in milliseconds) and calculate the elapsed time currentTime = glutGet(GLUT_ELAPSED_TIME); elapsedTime = currentTime - previousTime; /* Multiply the camera direction vector by constant speed then by the elapsed time (in seconds) and then move the camera */ SceneCamera.Move(cameraDirection * MOVEMENT_SPEED * (elapsedTime / 1000.0f)); // Requests to render a new frame (this will call my renderScene() once) glutPostRedisplay(); } void main(int argc, char **argv) { glutInit(&argc, argv); (...) glutDisplayFunc(renderScene); (...) // Setup the timer to be called one first time glutTimerFunc(TIMER_MILLISECONDS, processAnimationTimer, 0); // Read the current time since glutInit was called currentTime = glutGet(GLUT_ELAPSED_TIME); glutMainLoop(); } This implementation doesn't fell right. It works in the sense that helps the game speed to be constant dependent on the FPS. So that moving from point A to point B takes the same time no matter the high/low framerate. However, I believe I'm limiting the game framerate with this approach. Each frame will only be rendered when the time callback is called, that means the framerate will be roughly around TICKS_PER_SECOND frames per second. This doesn't feel right, you shouldn't limit your powerful hardware, it's wrong. It's my understanding though, that I still need to calculate the elapsedTime. Just because I'm telling GLUT to call the timer callback every TIMER_MILLISECONDS, it doesn't mean it will always do that on time. I'm not sure how can I fix this and to be completely honest, I have no idea what is the game loop in GLUT, you know, the while( game_is_running ) loop in Koen's article. But it's my understanding that GLUT is event-driven and that game loop starts when I call glutMainLoop() (which never returns), yes? I thought I could register an idle callback with glutIdleFunc() and use that as replacement of glutTimerFunc(), only rendering when necessary (instead of all the time as usual) but when I tested this with an empty callback (like void gameLoop() {}) and it was basically doing nothing, only a black screen, the CPU spiked to 25% and remained there until I killed the game and it went back to normal. So I don't think that's the path to follow. Using glutTimerFunc() is definitely not a good approach to perform all movements/animations based on that, as I'm limiting my game to a constant FPS, not cool. Or maybe I'm using it wrong and my implementation is not right? How exactly can I have a constant game speed with variable FPS? More exactly, how do I correctly implement Koen's Constant Game Speed with Maximum FPS solution (the fourth one on his article) with GLUT? Maybe this is not possible at all with GLUT? If not, what are my alternatives? What is the best approach to this problem (constant game speed) with GLUT? I originally posted this question on Stack Overflow before being pointed out about this site. The following is a different approach I tried after creating the question in SO, so I'm posting it here too. Another Approach: I've been experimenting and here's what I was able to achieve now. Instead of calculating the elapsed time on a timed function (which limits my game's framerate) I'm now doing it in renderScene(). Whenever changes to the scene happen I call glutPostRedisplay() (ie: camera moving, some object animation, etc...) which will make a call to renderScene(). I can use the elapsed time in this function to move my camera for instance. My code has now turned into this: int previousTime; int currentTime; int elapsedTime; void renderScene(void) { (...) // Setup the camera position and looking point SceneCamera.LookAt(); // Do all drawing below... (...) } void renderScene(void) { (...) // Get the time when the previous frame was rendered previousTime = currentTime; // Get the current time (in milliseconds) and calculate the elapsed time currentTime = glutGet(GLUT_ELAPSED_TIME); elapsedTime = currentTime - previousTime; /* Multiply the camera direction vector by constant speed then by the elapsed time (in seconds) and then move the camera */ SceneCamera.Move(cameraDirection * MOVEMENT_SPEED * (elapsedTime / 1000.0f)); // Setup the camera position and looking point SceneCamera.LookAt(); // All drawing code goes inside this function drawCompleteScene(); glutSwapBuffers(); /* Redraw the frame ONLY if the user is moving the camera (similar code will be needed to redraw the frame for other events) */ if(!IsTupleEmpty(cameraDirection)) { glutPostRedisplay(); } } void main(int argc, char **argv) { glutInit(&argc, argv); (...) glutDisplayFunc(renderScene); (...) currentTime = glutGet(GLUT_ELAPSED_TIME); glutMainLoop(); } Conclusion, it's working, or so it seems. If I don't move the camera, the CPU usage is low, nothing is being rendered (for testing purposes I only have a grid extending for 4000.0f, while zFar is set to 1000.0f). When I start moving the camera the scene starts redrawing itself. If I keep pressing the move keys, the CPU usage will increase; this is normal behavior. It drops back when I stop moving. Unless I'm missing something, it seems like a good approach for now. I did find this interesting article on iDevGames and this implementation is probably affected by the problem described on that article. What's your thoughts on that? Please note that I'm just doing this for fun, I have no intentions of creating some game to distribute or something like that, not in the near future at least. If I did, I would probably go with something else besides GLUT. But since I'm using GLUT, and other than the problem described on iDevGames, do you think this latest implementation is sufficient for GLUT? The only real issue I can think of right now is that I'll need to keep calling glutPostRedisplay() every time the scene changes something and keep calling it until there's nothing new to redraw. A little complexity added to the code for a better cause, I think. What do you think?

    Read the article

  • Configuration Manager Setting Causing Error PRJ0019

    - by Jeff Paterno
    Recently I ran into an issue with a project failing to build on an automated build server using CruiseControl. When I looked into the build log I saw that the Post-Build project was failing with the error message: "error PRJ0019: A tool returned an error code from "Performing Post-Build Event..." This was most frustrating especially since the solution was building without issue on my local development environment. The Post-Build project was a C++ project that basically called several batch files to unregister/register assemblies, copy resources and supporting files, and place other dependencies in the GAC. I decided to run each of the batch files manually to see if that would provide more information as to why this project was failing. This lead me to determine that the batch file that was placing assemblies in the GAC was the culprit and that it was failing to find a particular assembly. The missing assembly was the output of another project. The project that was not producing the expected output was another C++ project that called a batch file. This batch process was actually embedding resource files into an assembly and then copying the assembly to the expected location. The real confusion started when I looked back into my Subversion log and noted that nothing had changed in this project in more than 2 months! It was almost as if the project had stopped building altogether. But what would cause that?! The Configuration Manager, obviously! Checking the solution's Configuration Manager settings, I found that the project that was not producing any output was in fact not selected to be part of the build process when the "Any CPU" platform was selected. This was the problem! I had recently updated the CruiseControl configurations to force the solution to be built targeting the platform "Any CPU". As a result, the project that was at the root of the problem was not configured to be built and the post-build process was failing when it couldn't find what it needed.

    Read the article

  • Firefox 15 hangs with Ubuntu kernel update

    - by Marty
    I recently ran updates and it told me that in order to get those updates I had to update my kernel. I did that and also updated Firefox to 15. Since then Firefox hangs/gray screens sites I go to. This lasts anywhere from 5-10 seconds to 2-3 minutes. I have restarted Firefox with all add-ons disabled but it still did the same thing. I found a bug report on Launchpad that sounded like what was happening with me, but I haven't received any error codes, just the hanging/frozen screens. Also it seems that it ups my CPU making the rest of Ubuntu lag while Firefox is hung. I would guess the cause is a conflict between the updated kernel and the updated Firefox, but I'm still fairly new at Ubuntu and not sure where to go from here. Is there anything else to try? My Toshiba laptop specs are: Ubuntu 12.04 (32 bit) Linux 3.2.0-30-generic-pae #48-Ubuntu SMP Fri Aug 24 17:14:09 UTC 2012 i686 i686 i386 GNU/Linux Firefox 15.0.1 Intel® Pentium(R) Dual CPU T3400 @ 2.16GHz × 2 Mobile Intel® GM45 Express Chipset x86/MMX/SSE2 Thanks!

    Read the article

  • SQL SERVER – Using MAXDOP 1 for Single Processor Query – SQL in Sixty Seconds #008 – Video

    - by pinaldave
    Today’s SQL in Sixty Seconds video is inspired from my presentation at TechEd India 2012 on Speed up! – Parallel Processes and Unparalleled Performance. There are always special cases when it is about SQL Server. There are always few queries which gives optimal performance when they are executed on single processor and there are always queries which gives optimal performance when they are executed on multiple processors. I will be presenting the how to identify such queries as well what are the best practices related to the same. In this quick video I am going to demonstrate if the query is giving optimal performance when running on single CPU how one can restrict queries to single CPU by using hint OPTION (MAXDOP 1). More on Errors: Difference Temp Table and Table Variable – Effect of Transaction Effect of TRANSACTION on Local Variable – After ROLLBACK and After COMMIT Debate – Table Variables vs Temporary Tables – Quiz – Puzzle – 13 of 31 I encourage you to submit your ideas for SQL in Sixty Seconds. We will try to accommodate as many as we can. Reference: Pinal Dave (http://blog.sqlauthority.com) Filed under: Database, Pinal Dave, PostADay, SQL, SQL Authority, SQL in Sixty Seconds, SQL Query, SQL Scripts, SQL Server, SQL Tips and Tricks, SQLServer, T SQL, Video

    Read the article

  • 12.04.1 no audio through HDMI

    - by JoJo
    having a bit of an issue with getting audio to go through HDMI. Here are the base specs: OS: Ubuntu Desktop 12.04.1 x64 CPU: AMD A10-5800K 3.8G 4M FM2 R Mobo: MSI FM2-A75MA-E35 OS: Ubuntu 12.04.1 LTS Vid Card: (integrated on CPU) AMD Radeon HD 7660D HDMI sound works fine under Win7 (after mobo and vid drivers are installed), so it's not physically broken. Audio through the normal headphone jacks works fine under Ubuntu. Looking at the audio panel, there is no HDMI output at all. aplay -l also reports only: card 0: Generic [HD-Audio Generic], device 0: ALC887-VD Analog [ALC887-VD Analog] Subdevices: 1/1 Subdevice #0: subdevice #0 In additional drivers there are two versions: ATI/AMD PROPRIETARY FGLRX GRAPHICS DRIVER ATI/AMD PROPRIETARY FGLRX GRAPHICS DRIVER (post-release update) The first installs, but problem persists. I do get more resolutions to pick from. Second version does not, reporting it failed installation and to find details at: /var/log/jockey.log Looked at the log, and it's insanely long, if necessary I can get it to you guys. Did more research and some had luck by manually installing the drivers, so tried to give that a shot by following this: https://help.ubuntu.com/community/BinaryDriverHowto/ATI#Manually_installing_Catalyst_12.6 starting at 3.1 Manually installing Catalyst 12.6. I immediately had 2 issues, (1) the AMD website does not provide any drivers for Linux, and (2) the following command did not work: sudo sh amd-driver-installer-12-6-x86.x86_64.run --buildpkg Ubuntu/precise sh: 0: Can't open amd-driver-installer-12-6-x86.x86_64.run Some other posts stated to update "alsa-drivers", but that also did not work as install command for the new version of them did not work. I forget the exact issue, but similar to above, cannot open / cannot find. Any help would be greatly appreciated!

    Read the article

  • Effective way to check if an Entity/Player enters a region/trigger

    - by Chris
    I was wondering how multiplayer games detect if you enter a special region. Let's assume there is a huge map that is so big that simply checking it would become a huge performance issue. I've seen bukkit (a modding API for Minecraft servers) firing an Event on every single move. I don't think that larger games do the same because even if you have only a few coordinates you are interested in, you have to loop through a few trigger zone to see if the player is inside your region - for every player. This seems like an extremely CPU-intense operation to me even though I've never developed something like that. Is there a special algorithm that is used by larger games to accomplish this? The only thing I could imagine is to split up the world into multiple parts and to register the event not on the movement itself but on all the parts that are covered by your area and only check for areas that are registered in the current part. And another thing I would like to know: How could you detect when someone must have entered a trigger but you never saw him directly in it since his client only sent you an move packet shortly before entering and after leaving the trigger area. Drawing a line and calculate all colliding parts seems rather CPU intensive if you have to perform it every time.

    Read the article

  • Can windows XP be better than any Ubuntu (and Linux) distro for an old PC?

    - by Robert Vila
    The old laptop is a Toshiba 1800-100: CPU: Intel Celeron 800h Ram 128 MB (works ok) HDD: 15GB (works ok) Graphics adapter: Integrated 64-bit AGP graphics accelerator, BitBIT, 3D graphic acceleration, 8 MB Video RAM Only WindowsXP is installed, and works ok: it can be used, but it is slow (and hateful). I thought that I could improve performance (and its look) easily, since it is an old PC (drivers and everything known for years...) by installing a light Linux distro. So, I decided to install a light or customized Ubuntu distro, or Ubuntu/Debian derivative, but haven't been successful with any; not even booting LiveCDs: not even AntiX, not even Puppy. Lubuntu wiki says that it won't work because the last to releases need more ram (and some blogs say much more cpu -even core duo for new Lubuntu!-), let alone Xubuntu. The problems I have faced are: 1.There are thousands of pages talking about the same 10/15 lightweight distros, and saying more or less the same things, but NONE talks about a simple thing as to how should the RAM/swap-partition proportion be for this kind of installations. NONE! 2.Loading the LiveCD I have tried several different boot options (don't understand much about this and there's ALWAYS a line of explanation missing) and never receive error messages. Booting just stops at different stages but often seems to stop just when the X server is going to start. I am able to boot to command line. 3.I ignore whether the problem is ram size or a problem with the graphics driver (which surprises me because it is a well known brand and line of computers). So I don't know if doing a partition with a swap partition would help booting the LiveCD. 4.I would like to try the graphical interface with the LiveCD before installing. If doing the swap partition for this purpose would help. How can I do the partition? I tried to use Boot Rescue CD, but it advises me against continuing forward. I would appreciate any ideas as regards these questions. Thank you

    Read the article

  • Data binding in web UI frameworks, what's the deal?

    - by c-smile
    I believe that most of modern Web frameworks that pretend to be MVC ones also has a notion of data binding in one form or another. Examples: AngularJS, EmberJS, KnockoutJS, etc. I am assuming that "data binding" is a declarative definition (oxymoron, no?) of live link between data (a.k.a. model) and its representation (a.k.a. view). With some transformers in between (a.k.a. controllers). I understand why declarativeness is kind of appealing but also understand that as usual it comes with the price. In particular: 1. Live binding is quite heavy, either with dirty watch (high CPU consumption) or with Object.observe() (high memory consumption with high CPU load in some scenarios). 2. There is a "frame" part in the framework word, means there are some boundaries/limits that can be hard to overcome if you need slightly more than it was designed for. Quite usual time split: 90% of features are made in 10% of project time. But 10% rest take 90% of project time. I suspect (a.k.a. educated guess) that those MVC things are not helping to implement more functionality in less time... If so their usage motivation is not quite clear. As an example: last week wanted to find virtual list idea/solution. Found one in vanilla JavaScript that is 120 LOC. Implementation of the same but in AngualrJS is about 420 LOC. Most of the code there seems like a fight with the framework itself... So is my question: what benefits that MVC stuff or data binding give us? Is it just a buzzword popular among project managers or they give us something useful. If later one then what exactly?

    Read the article

  • ATI Catalyst driver 12.8 is not using hardware acceleration on Precise

    - by Jack Wright
    I've been using Ubuntu and ATI Catalyst for years. On my clean install of Ubuntu 12.04 I've noticed that Catalyst 12.6 and then 12.8 are not actually using my HD5750 GPU for hardware acceleration - high CPU usage, zero GPU load. Everything installed correctly with no hassles, fglrxinfo and vainfo are correct as per this HowTo for Precise. I have an Ubuntu 10.04 with Catalyst 12.6 installation on the same hardware which does use the GPU - low CPU usage, high GPU load when transcodeing video files or playing video content. The VA-API drivers are not installed on the 10.04 build. They are not mentioned in this HowTo for Lucid. fgl_glxgears frame rates on Precise are a fifth of the rates on Lucid. LUCID jw@Kworld:~$ fgl_glxgears Using GLX_SGIX_pbuffer 16867 frames in 5.0 seconds = 3373.400 FPS 12523 frames in 5.0 seconds = 2504.600 FPS 13763 frames in 5.0 seconds = 2752.600 FPS PRECISE jw@NewWorld12:~$ fgl_glxgears Using GLX_SGIX_pbuffer 12905 frames in 5.0 seconds = 2581.000 FPS 3230 frames in 5.0 seconds = 646.000 FPS 517 frames in 5.0 seconds = 103.400 FPS 518 frames in 5.0 seconds = 103.600 FPS 6489 frames in 5.0 seconds = 1297.800 FPS This is glxgears running in fullscreen. In Lucid (10.04) I can't see the gears, they are spinning so fast, but in Precise (12.04) they are really sluggish. Has anyone else noticed a problem like this? Cheers, Jack.

    Read the article

  • Why Are Minimized Programs Often Slow to Open Again?

    - by Jason Fitzpatrick
    It seems particularly counterintuitive: you minimize an application because you plan on returning to it later and wish to skip shutting the application down and restarting it later, but sometimes maximizing it takes even longer than launching it fresh. What gives? Today’s Question & Answer session comes to us courtesy of SuperUser—a subdivision of Stack Exchange, a community-driven grouping of Q&A web sites. The Question SuperUser reader Bart wants to know why he’s not saving any time with application minimization: I’m working in Photoshop CS6 and multiple browsers a lot. I’m not using them all at once, so sometimes some applications are minimized to taskbar for hours or days. The problem is, when I try to maximize them from the taskbar – it sometimes takes longer than starting them! Especially Photoshop feels really weird for many seconds after finally showing up, it’s slow, unresponsive and even sometimes totally freezes for minute or two. It’s not a hardware problem as it’s been like that since always on all on my PCs. Would I also notice it after upgrading my HDD to SDD and adding RAM (my main PC holds 4 GB currently)? Could guys with powerful pcs / macs tell me – does it also happen to you? I guess OSes somehow “focus” on active software and move all the resources away from the ones that run, but are not used. Is it possible to somehow set RAM / CPU / HDD priorities or something, for let’s say, Photoshop, so it won’t slow down after long period of inactivity? So what is the deal? Why does he find himself waiting to maximize a minimized app? The Answer SuperUser contributor Allquixotic explains why: Summary The immediate problem is that the programs that you have minimized are being paged out to the “page file” on your hard disk. This symptom can be improved by installing a Solid State Disk (SSD), adding more RAM to your system, reducing the number of programs you have open, or upgrading to a newer system architecture (for instance, Ivy Bridge or Haswell). Out of these options, adding more RAM is generally the most effective solution. Explanation The default behavior of Windows is to give active applications priority over inactive applications for having a spot in RAM. When there’s significant memory pressure (meaning the system doesn’t have a lot of free RAM if it were to let every program have all the RAM it wants), it starts putting minimized programs into the page file, which means it writes out their contents from RAM to disk, and then makes that area of RAM free. That free RAM helps programs you’re actively using — say, your web browser — run faster, because if they need to claim a new segment of RAM (like when you open a new tab), they can do so. This “free” RAM is also used as page cache, which means that when active programs attempt to read data on your hard disk, that data might be cached in RAM, which prevents your hard disk from being accessed to get that data. By using the majority of your RAM for page cache, and swapping out unused programs to disk, Windows is trying to improve responsiveness of the program(s) you are actively using, by making RAM available to them, and caching the files they access in RAM instead of the hard disk. The downside of this behavior is that minimized programs can take a while to have their contents copied from the page file, on disk, back into RAM. The time increases the larger the program’s footprint in memory. This is why you experience that delay when maximizing Photoshop. RAM is many times faster than a hard disk (depending on the specific hardware, it can be up to several orders of magnitude). An SSD is considerably faster than a hard disk, but it is still slower than RAM by orders of magnitude. Having your page file on an SSD will help, but it will also wear out the SSD more quickly than usual if your page file is heavily utilized due to RAM pressure. Remedies Here is an explanation of the available remedies, and their general effectiveness: Installing more RAM: This is the recommended path. If your system does not support more RAM than you already have installed, you will need to upgrade more of your system: possibly your motherboard, CPU, chassis, power supply, etc. depending on how old it is. If it’s a laptop, chances are you’ll have to buy an entire new laptop that supports more installed RAM. When you install more RAM, you reduce memory pressure, which reduces use of the page file, which is a good thing all around. You also make available more RAM for page cache, which will make all programs that access the hard disk run faster. As of Q4 2013, my personal recommendation is that you have at least 8 GB of RAM for a desktop or laptop whose purpose is anything more complex than web browsing and email. That means photo editing, video editing/viewing, playing computer games, audio editing or recording, programming / development, etc. all should have at least 8 GB of RAM, if not more. Run fewer programs at a time: This will only work if the programs you are running do not use a lot of memory on their own. Unfortunately, Adobe Creative Suite products such as Photoshop CS6 are known for using an enormous amount of memory. This also limits your multitasking ability. It’s a temporary, free remedy, but it can be an inconvenience to close down your web browser or Word every time you start Photoshop, for instance. This also wouldn’t stop Photoshop from being swapped when minimizing it, so it really isn’t a very effective solution. It only helps in some specific situations. Install an SSD: If your page file is on an SSD, the SSD’s improved speed compared to a hard disk will result in generally improved performance when the page file has to be read from or written to. Be aware that SSDs are not designed to withstand a very frequent and constant random stream of writes; they can only be written over a limited number of times before they start to break down. Heavy use of a page file is not a particularly good workload for an SSD. You should install an SSD in combination with a large amount of RAM if you want maximum performance while preserving the longevity of the SSD. Use a newer system architecture: Depending on the age of your system, you may be using an out of date system architecture. The “system architecture” is generally defined as the “generation” (think generations like children, parents, grandparents, etc.) of the motherboard and CPU. Newer generations generally support faster I/O (input/output), better memory bandwidth, lower latency, and less contention over shared resources, instead providing dedicated links between components. For example, starting with the “Nehalem” generation (around 2009), the Front-Side Bus (FSB) was eliminated, which removed a common bottleneck, because almost all system components had to share the same FSB for transmitting data. This was replaced with a “point to point” architecture, meaning that each component gets its own dedicated “lane” to the CPU, which continues to be improved every few years with new generations. You will generally see a more significant improvement in overall system performance depending on the “gap” between your computer’s architecture and the latest one available. For example, a Pentium 4 architecture from 2004 is going to see a much more significant improvement upgrading to “Haswell” (the latest as of Q4 2013) than a “Sandy Bridge” architecture from ~2010. Links Related questions: How to reduce disk thrashing (paging)? Windows Swap (Page File): Enable or Disable? Also, just in case you’re considering it, you really shouldn’t disable the page file, as this will only make matters worse; see here. And, in case you needed extra convincing to leave the Windows Page File alone, see here and here. Have something to add to the explanation? Sound off in the the comments. Want to read more answers from other tech-savvy Stack Exchange users? Check out the full discussion thread here.     

    Read the article

< Previous Page | 90 91 92 93 94 95 96 97 98 99 100 101  | Next Page >