Search Results

Search found 454 results on 19 pages for 'optimizations'.

Page 12/19 | < Previous Page | 8 9 10 11 12 13 14 15 16 17 18 19  | Next Page >

  • Amazon EC2: Instances, IPs and a wordpress blog (LAMP)

    - by JustinXXVII
    I had a link to my blog posted on Reddit yesterday and MySQL crashed on my EC2 Micro instance. I know I didn't have that many visitors because I used a marketing link that tracks hits. The link got 167 hits over the course of the last 18 hours, and MySQL crashed twice. So anyway, 167 visits is not a lot, so I've done some short term optimizations like restricting the number of Apache threads to limit the MySQL calls. I also set up WP Super Cache to serve static content. Soon I'm going to offload all of my images to S3 or CloudFront. So this leads me to my question. If this doesn't seem to help, and if i have another traffic "spike", how do AMIs work when you have a MySQL database? I think I understand that if you have more than one instance and assign the same Elastic IP to both of them, the incoming traffic gets distributed among both. But what happens when the MySQL database gets updated on one of the instances? I just need to wrap my mind around what happens when I create an AMI and then launch a new instance to help with traffic. Thanks for your suggestions.

    Read the article

  • My system is always disk-bound (the disk light is always on). Why is this?

    - by Scoobie
    I have been given a laptop by the good folks at my company on which to do my work (Java development). I usually use eclipse as my primary development platform. The laptop is a Dell D830 and runs Windows 7 - 32 bit. Although the processor supports a 64 bit instruction-set, licensing limits me to running the 32 bit OS. The HDD is a WD1600BEVT (Western Digital). I have noticed that my disk is always very slow. Windows start up is usually pretty quick, however as soon as I log on, my disk light stays on and usually, the laptop takes about 4 minutes (after logging in -- immediately upon getting the prompt to press Ctrl + Alt + Del to log in) before it's usable. Questions: Is this expected behavior? What can I do to examine the disk and determine the cause of the problem? What can I do to improve my disk's performance? Any optimizations you may be able to suggest? Other Questions: Some have suggested running Process Monitor (from sysinternals), but how would i get the log since start up? Instead of trying to fix this myself, should I simply push this onto the system administrator? Thanks all.

    Read the article

  • value types in the vm

    - by john.rose
    value types in the vm p.p1 {margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px Times} p.p2 {margin: 0.0px 0.0px 14.0px 0.0px; font: 14.0px Times} p.p3 {margin: 0.0px 0.0px 12.0px 0.0px; font: 14.0px Times} p.p4 {margin: 0.0px 0.0px 15.0px 0.0px; font: 14.0px Times} p.p5 {margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px Courier} p.p6 {margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px Courier; min-height: 17.0px} p.p7 {margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px Times; min-height: 18.0px} p.p8 {margin: 0.0px 0.0px 0.0px 36.0px; text-indent: -36.0px; font: 14.0px Times; min-height: 18.0px} p.p9 {margin: 0.0px 0.0px 12.0px 0.0px; font: 14.0px Times; min-height: 18.0px} p.p10 {margin: 0.0px 0.0px 12.0px 0.0px; font: 14.0px Times; color: #000000} li.li1 {margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px Times} li.li7 {margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px Times; min-height: 18.0px} span.s1 {font: 14.0px Courier} span.s2 {color: #000000} span.s3 {font: 14.0px Courier; color: #000000} ol.ol1 {list-style-type: decimal} Or, enduring values for a changing world. Introduction A value type is a data type which, generally speaking, is designed for being passed by value in and out of methods, and stored by value in data structures. The only value types which the Java language directly supports are the eight primitive types. Java indirectly and approximately supports value types, if they are implemented in terms of classes. For example, both Integer and String may be viewed as value types, especially if their usage is restricted to avoid operations appropriate to Object. In this note, we propose a definition of value types in terms of a design pattern for Java classes, accompanied by a set of usage restrictions. We also sketch the relation of such value types to tuple types (which are a JVM-level notion), and point out JVM optimizations that can apply to value types. This note is a thought experiment to extend the JVM’s performance model in support of value types. The demonstration has two phases.  Initially the extension can simply use design patterns, within the current bytecode architecture, and in today’s Java language. But if the performance model is to be realized in practice, it will probably require new JVM bytecode features, changes to the Java language, or both.  We will look at a few possibilities for these new features. An Axiom of Value In the context of the JVM, a value type is a data type equipped with construction, assignment, and equality operations, and a set of typed components, such that, whenever two variables of the value type produce equal corresponding values for their components, the values of the two variables cannot be distinguished by any JVM operation. Here are some corollaries: A value type is immutable, since otherwise a copy could be constructed and the original could be modified in one of its components, allowing the copies to be distinguished. Changing the component of a value type requires construction of a new value. The equals and hashCode operations are strictly component-wise. If a value type is represented by a JVM reference, that reference cannot be successfully synchronized on, and cannot be usefully compared for reference equality. A value type can be viewed in terms of what it doesn’t do. We can say that a value type omits all value-unsafe operations, which could violate the constraints on value types.  These operations, which are ordinarily allowed for Java object types, are pointer equality comparison (the acmp instruction), synchronization (the monitor instructions), all the wait and notify methods of class Object, and non-trivial finalize methods. The clone method is also value-unsafe, although for value types it could be treated as the identity function. Finally, and most importantly, any side effect on an object (however visible) also counts as an value-unsafe operation. A value type may have methods, but such methods must not change the components of the value. It is reasonable and useful to define methods like toString, equals, and hashCode on value types, and also methods which are specifically valuable to users of the value type. Representations of Value Value types have two natural representations in the JVM, unboxed and boxed. An unboxed value consists of the components, as simple variables. For example, the complex number x=(1+2i), in rectangular coordinate form, may be represented in unboxed form by the following pair of variables: /*Complex x = Complex.valueOf(1.0, 2.0):*/ double x_re = 1.0, x_im = 2.0; These variables might be locals, parameters, or fields. Their association as components of a single value is not defined to the JVM. Here is a sample computation which computes the norm of the difference between two complex numbers: double distance(/*Complex x:*/ double x_re, double x_im,         /*Complex y:*/ double y_re, double y_im) {     /*Complex z = x.minus(y):*/     double z_re = x_re - y_re, z_im = x_im - y_im;     /*return z.abs():*/     return Math.sqrt(z_re*z_re + z_im*z_im); } A boxed representation groups component values under a single object reference. The reference is to a ‘wrapper class’ that carries the component values in its fields. (A primitive type can naturally be equated with a trivial value type with just one component of that type. In that view, the wrapper class Integer can serve as a boxed representation of value type int.) The unboxed representation of complex numbers is practical for many uses, but it fails to cover several major use cases: return values, array elements, and generic APIs. The two components of a complex number cannot be directly returned from a Java function, since Java does not support multiple return values. The same story applies to array elements: Java has no ’array of structs’ feature. (Double-length arrays are a possible workaround for complex numbers, but not for value types with heterogeneous components.) By generic APIs I mean both those which use generic types, like Arrays.asList and those which have special case support for primitive types, like String.valueOf and PrintStream.println. Those APIs do not support unboxed values, and offer some problems to boxed values. Any ’real’ JVM type should have a story for returns, arrays, and API interoperability. The basic problem here is that value types fall between primitive types and object types. Value types are clearly more complex than primitive types, and object types are slightly too complicated. Objects are a little bit dangerous to use as value carriers, since object references can be compared for pointer equality, and can be synchronized on. Also, as many Java programmers have observed, there is often a performance cost to using wrapper objects, even on modern JVMs. Even so, wrapper classes are a good starting point for talking about value types. If there were a set of structural rules and restrictions which would prevent value-unsafe operations on value types, wrapper classes would provide a good notation for defining value types. This note attempts to define such rules and restrictions. Let’s Start Coding Now it is time to look at some real code. Here is a definition, written in Java, of a complex number value type. @ValueSafe public final class Complex implements java.io.Serializable {     // immutable component structure:     public final double re, im;     private Complex(double re, double im) {         this.re = re; this.im = im;     }     // interoperability methods:     public String toString() { return "Complex("+re+","+im+")"; }     public List<Double> asList() { return Arrays.asList(re, im); }     public boolean equals(Complex c) {         return re == c.re && im == c.im;     }     public boolean equals(@ValueSafe Object x) {         return x instanceof Complex && equals((Complex) x);     }     public int hashCode() {         return 31*Double.valueOf(re).hashCode()                 + Double.valueOf(im).hashCode();     }     // factory methods:     public static Complex valueOf(double re, double im) {         return new Complex(re, im);     }     public Complex changeRe(double re2) { return valueOf(re2, im); }     public Complex changeIm(double im2) { return valueOf(re, im2); }     public static Complex cast(@ValueSafe Object x) {         return x == null ? ZERO : (Complex) x;     }     // utility methods and constants:     public Complex plus(Complex c)  { return new Complex(re+c.re, im+c.im); }     public Complex minus(Complex c) { return new Complex(re-c.re, im-c.im); }     public double abs() { return Math.sqrt(re*re + im*im); }     public static final Complex PI = valueOf(Math.PI, 0.0);     public static final Complex ZERO = valueOf(0.0, 0.0); } This is not a minimal definition, because it includes some utility methods and other optional parts.  The essential elements are as follows: The class is marked as a value type with an annotation. The class is final, because it does not make sense to create subclasses of value types. The fields of the class are all non-private and final.  (I.e., the type is immutable and structurally transparent.) From the supertype Object, all public non-final methods are overridden. The constructor is private. Beyond these bare essentials, we can observe the following features in this example, which are likely to be typical of all value types: One or more factory methods are responsible for value creation, including a component-wise valueOf method. There are utility methods for complex arithmetic and instance creation, such as plus and changeIm. There are static utility constants, such as PI. The type is serializable, using the default mechanisms. There are methods for converting to and from dynamically typed references, such as asList and cast. The Rules In order to use value types properly, the programmer must avoid value-unsafe operations.  A helpful Java compiler should issue errors (or at least warnings) for code which provably applies value-unsafe operations, and should issue warnings for code which might be correct but does not provably avoid value-unsafe operations.  No such compilers exist today, but to simplify our account here, we will pretend that they do exist. A value-safe type is any class, interface, or type parameter marked with the @ValueSafe annotation, or any subtype of a value-safe type.  If a value-safe class is marked final, it is in fact a value type.  All other value-safe classes must be abstract.  The non-static fields of a value class must be non-public and final, and all its constructors must be private. Under the above rules, a standard interface could be helpful to define value types like Complex.  Here is an example: @ValueSafe public interface ValueType extends java.io.Serializable {     // All methods listed here must get redefined.     // Definitions must be value-safe, which means     // they may depend on component values only.     List<? extends Object> asList();     int hashCode();     boolean equals(@ValueSafe Object c);     String toString(); } //@ValueSafe inherited from supertype: public final class Complex implements ValueType { … The main advantage of such a conventional interface is that (unlike an annotation) it is reified in the runtime type system.  It could appear as an element type or parameter bound, for facilities which are designed to work on value types only.  More broadly, it might assist the JVM to perform dynamic enforcement of the rules for value types. Besides types, the annotation @ValueSafe can mark fields, parameters, local variables, and methods.  (This is redundant when the type is also value-safe, but may be useful when the type is Object or another supertype of a value type.)  Working forward from these annotations, an expression E is defined as value-safe if it satisfies one or more of the following: The type of E is a value-safe type. E names a field, parameter, or local variable whose declaration is marked @ValueSafe. E is a call to a method whose declaration is marked @ValueSafe. E is an assignment to a value-safe variable, field reference, or array reference. E is a cast to a value-safe type from a value-safe expression. E is a conditional expression E0 ? E1 : E2, and both E1 and E2 are value-safe. Assignments to value-safe expressions and initializations of value-safe names must take their values from value-safe expressions. A value-safe expression may not be the subject of a value-unsafe operation.  In particular, it cannot be synchronized on, nor can it be compared with the “==” operator, not even with a null or with another value-safe type. In a program where all of these rules are followed, no value-type value will be subject to a value-unsafe operation.  Thus, the prime axiom of value types will be satisfied, that no two value type will be distinguishable as long as their component values are equal. More Code To illustrate these rules, here are some usage examples for Complex: Complex pi = Complex.valueOf(Math.PI, 0); Complex zero = pi.changeRe(0);  //zero = pi; zero.re = 0; ValueType vtype = pi; @SuppressWarnings("value-unsafe")   Object obj = pi; @ValueSafe Object obj2 = pi; obj2 = new Object();  // ok List<Complex> clist = new ArrayList<Complex>(); clist.add(pi);  // (ok assuming List.add param is @ValueSafe) List<ValueType> vlist = new ArrayList<ValueType>(); vlist.add(pi);  // (ok) List<Object> olist = new ArrayList<Object>(); olist.add(pi);  // warning: "value-unsafe" boolean z = pi.equals(zero); boolean z1 = (pi == zero);  // error: reference comparison on value type boolean z2 = (pi == null);  // error: reference comparison on value type boolean z3 = (pi == obj2);  // error: reference comparison on value type synchronized (pi) { }  // error: synch of value, unpredictable result synchronized (obj2) { }  // unpredictable result Complex qq = pi; qq = null;  // possible NPE; warning: “null-unsafe" qq = (Complex) obj;  // warning: “null-unsafe" qq = Complex.cast(obj);  // OK @SuppressWarnings("null-unsafe")   Complex empty = null;  // possible NPE qq = empty;  // possible NPE (null pollution) The Payoffs It follows from this that either the JVM or the java compiler can replace boxed value-type values with unboxed ones, without affecting normal computations.  Fields and variables of value types can be split into their unboxed components.  Non-static methods on value types can be transformed into static methods which take the components as value parameters. Some common questions arise around this point in any discussion of value types. Why burden the programmer with all these extra rules?  Why not detect programs automagically and perform unboxing transparently?  The answer is that it is easy to break the rules accidently unless they are agreed to by the programmer and enforced.  Automatic unboxing optimizations are tantalizing but (so far) unreachable ideal.  In the current state of the art, it is possible exhibit benchmarks in which automatic unboxing provides the desired effects, but it is not possible to provide a JVM with a performance model that assures the programmer when unboxing will occur.  This is why I’m writing this note, to enlist help from, and provide assurances to, the programmer.  Basically, I’m shooting for a good set of user-supplied “pragmas” to frame the desired optimization. Again, the important thing is that the unboxing must be done reliably, or else programmers will have no reason to work with the extra complexity of the value-safety rules.  There must be a reasonably stable performance model, wherein using a value type has approximately the same performance characteristics as writing the unboxed components as separate Java variables. There are some rough corners to the present scheme.  Since Java fields and array elements are initialized to null, value-type computations which incorporate uninitialized variables can produce null pointer exceptions.  One workaround for this is to require such variables to be null-tested, and the result replaced with a suitable all-zero value of the value type.  That is what the “cast” method does above. Generically typed APIs like List<T> will continue to manipulate boxed values always, at least until we figure out how to do reification of generic type instances.  Use of such APIs will elicit warnings until their type parameters (and/or relevant members) are annotated or typed as value-safe.  Retrofitting List<T> is likely to expose flaws in the present scheme, which we will need to engineer around.  Here are a couple of first approaches: public interface java.util.List<@ValueSafe T> extends Collection<T> { … public interface java.util.List<T extends Object|ValueType> extends Collection<T> { … (The second approach would require disjunctive types, in which value-safety is “contagious” from the constituent types.) With more transformations, the return value types of methods can also be unboxed.  This may require significant bytecode-level transformations, and would work best in the presence of a bytecode representation for multiple value groups, which I have proposed elsewhere under the title “Tuples in the VM”. But for starters, the JVM can apply this transformation under the covers, to internally compiled methods.  This would give a way to express multiple return values and structured return values, which is a significant pain-point for Java programmers, especially those who work with low-level structure types favored by modern vector and graphics processors.  The lack of multiple return values has a strong distorting effect on many Java APIs. Even if the JVM fails to unbox a value, there is still potential benefit to the value type.  Clustered computing systems something have copy operations (serialization or something similar) which apply implicitly to command operands.  When copying JVM objects, it is extremely helpful to know when an object’s identity is important or not.  If an object reference is a copied operand, the system may have to create a proxy handle which points back to the original object, so that side effects are visible.  Proxies must be managed carefully, and this can be expensive.  On the other hand, value types are exactly those types which a JVM can “copy and forget” with no downside. Array types are crucial to bulk data interfaces.  (As data sizes and rates increase, bulk data becomes more important than scalar data, so arrays are definitely accompanying us into the future of computing.)  Value types are very helpful for adding structure to bulk data, so a successful value type mechanism will make it easier for us to express richer forms of bulk data. Unboxing arrays (i.e., arrays containing unboxed values) will provide better cache and memory density, and more direct data movement within clustered or heterogeneous computing systems.  They require the deepest transformations, relative to today’s JVM.  There is an impedance mismatch between value-type arrays and Java’s covariant array typing, so compromises will need to be struck with existing Java semantics.  It is probably worth the effort, since arrays of unboxed value types are inherently more memory-efficient than standard Java arrays, which rely on dependent pointer chains. It may be sufficient to extend the “value-safe” concept to array declarations, and allow low-level transformations to change value-safe array declarations from the standard boxed form into an unboxed tuple-based form.  Such value-safe arrays would not be convertible to Object[] arrays.  Certain connection points, such as Arrays.copyOf and System.arraycopy might need additional input/output combinations, to allow smooth conversion between arrays with boxed and unboxed elements. Alternatively, the correct solution may have to wait until we have enough reification of generic types, and enough operator overloading, to enable an overhaul of Java arrays. Implicit Method Definitions The example of class Complex above may be unattractively complex.  I believe most or all of the elements of the example class are required by the logic of value types. If this is true, a programmer who writes a value type will have to write lots of error-prone boilerplate code.  On the other hand, I think nearly all of the code (except for the domain-specific parts like plus and minus) can be implicitly generated. Java has a rule for implicitly defining a class’s constructor, if no it defines no constructors explicitly.  Likewise, there are rules for providing default access modifiers for interface members.  Because of the highly regular structure of value types, it might be reasonable to perform similar implicit transformations on value types.  Here’s an example of a “highly implicit” definition of a complex number type: public class Complex implements ValueType {  // implicitly final     public double re, im;  // implicitly public final     //implicit methods are defined elementwise from te fields:     //  toString, asList, equals(2), hashCode, valueOf, cast     //optionally, explicit methods (plus, abs, etc.) would go here } In other words, with the right defaults, a simple value type definition can be a one-liner.  The observant reader will have noticed the similarities (and suitable differences) between the explicit methods above and the corresponding methods for List<T>. Another way to abbreviate such a class would be to make an annotation the primary trigger of the functionality, and to add the interface(s) implicitly: public @ValueType class Complex { … // implicitly final, implements ValueType (But to me it seems better to communicate the “magic” via an interface, even if it is rooted in an annotation.) Implicitly Defined Value Types So far we have been working with nominal value types, which is to say that the sequence of typed components is associated with a name and additional methods that convey the intention of the programmer.  A simple ordered pair of floating point numbers can be variously interpreted as (to name a few possibilities) a rectangular or polar complex number or Cartesian point.  The name and the methods convey the intended meaning. But what if we need a truly simple ordered pair of floating point numbers, without any further conceptual baggage?  Perhaps we are writing a method (like “divideAndRemainder”) which naturally returns a pair of numbers instead of a single number.  Wrapping the pair of numbers in a nominal type (like “QuotientAndRemainder”) makes as little sense as wrapping a single return value in a nominal type (like “Quotient”).  What we need here are structural value types commonly known as tuples. For the present discussion, let us assign a conventional, JVM-friendly name to tuples, roughly as follows: public class java.lang.tuple.$DD extends java.lang.tuple.Tuple {      double $1, $2; } Here the component names are fixed and all the required methods are defined implicitly.  The supertype is an abstract class which has suitable shared declarations.  The name itself mentions a JVM-style method parameter descriptor, which may be “cracked” to determine the number and types of the component fields. The odd thing about such a tuple type (and structural types in general) is it must be instantiated lazily, in response to linkage requests from one or more classes that need it.  The JVM and/or its class loaders must be prepared to spin a tuple type on demand, given a simple name reference, $xyz, where the xyz is cracked into a series of component types.  (Specifics of naming and name mangling need some tasteful engineering.) Tuples also seem to demand, even more than nominal types, some support from the language.  (This is probably because notations for non-nominal types work best as combinations of punctuation and type names, rather than named constructors like Function3 or Tuple2.)  At a minimum, languages with tuples usually (I think) have some sort of simple bracket notation for creating tuples, and a corresponding pattern-matching syntax (or “destructuring bind”) for taking tuples apart, at least when they are parameter lists.  Designing such a syntax is no simple thing, because it ought to play well with nominal value types, and also with pre-existing Java features, such as method parameter lists, implicit conversions, generic types, and reflection.  That is a task for another day. Other Use Cases Besides complex numbers and simple tuples there are many use cases for value types.  Many tuple-like types have natural value-type representations. These include rational numbers, point locations and pixel colors, and various kinds of dates and addresses. Other types have a variable-length ‘tail’ of internal values. The most common example of this is String, which is (mathematically) a sequence of UTF-16 character values. Similarly, bit vectors, multiple-precision numbers, and polynomials are composed of sequences of values. Such types include, in their representation, a reference to a variable-sized data structure (often an array) which (somehow) represents the sequence of values. The value type may also include ’header’ information. Variable-sized values often have a length distribution which favors short lengths. In that case, the design of the value type can make the first few values in the sequence be direct ’header’ fields of the value type. In the common case where the header is enough to represent the whole value, the tail can be a shared null value, or even just a null reference. Note that the tail need not be an immutable object, as long as the header type encapsulates it well enough. This is the case with String, where the tail is a mutable (but never mutated) character array. Field types and their order must be a globally visible part of the API.  The structure of the value type must be transparent enough to have a globally consistent unboxed representation, so that all callers and callees agree about the type and order of components  that appear as parameters, return types, and array elements.  This is a trade-off between efficiency and encapsulation, which is forced on us when we remove an indirection enjoyed by boxed representations.  A JVM-only transformation would not care about such visibility, but a bytecode transformation would need to take care that (say) the components of complex numbers would not get swapped after a redefinition of Complex and a partial recompile.  Perhaps constant pool references to value types need to declare the field order as assumed by each API user. This brings up the delicate status of private fields in a value type.  It must always be possible to load, store, and copy value types as coordinated groups, and the JVM performs those movements by moving individual scalar values between locals and stack.  If a component field is not public, what is to prevent hostile code from plucking it out of the tuple using a rogue aload or astore instruction?  Nothing but the verifier, so we may need to give it more smarts, so that it treats value types as inseparable groups of stack slots or locals (something like long or double). My initial thought was to make the fields always public, which would make the security problem moot.  But public is not always the right answer; consider the case of String, where the underlying mutable character array must be encapsulated to prevent security holes.  I believe we can win back both sides of the tradeoff, by training the verifier never to split up the components in an unboxed value.  Just as the verifier encapsulates the two halves of a 64-bit primitive, it can encapsulate the the header and body of an unboxed String, so that no code other than that of class String itself can take apart the values. Similar to String, we could build an efficient multi-precision decimal type along these lines: public final class DecimalValue extends ValueType {     protected final long header;     protected private final BigInteger digits;     public DecimalValue valueOf(int value, int scale) {         assert(scale >= 0);         return new DecimalValue(((long)value << 32) + scale, null);     }     public DecimalValue valueOf(long value, int scale) {         if (value == (int) value)             return valueOf((int)value, scale);         return new DecimalValue(-scale, new BigInteger(value));     } } Values of this type would be passed between methods as two machine words. Small values (those with a significand which fits into 32 bits) would be represented without any heap data at all, unless the DecimalValue itself were boxed. (Note the tension between encapsulation and unboxing in this case.  It would be better if the header and digits fields were private, but depending on where the unboxing information must “leak”, it is probably safer to make a public revelation of the internal structure.) Note that, although an array of Complex can be faked with a double-length array of double, there is no easy way to fake an array of unboxed DecimalValues.  (Either an array of boxed values or a transposed pair of homogeneous arrays would be reasonable fallbacks, in a current JVM.)  Getting the full benefit of unboxing and arrays will require some new JVM magic. Although the JVM emphasizes portability, system dependent code will benefit from using machine-level types larger than 64 bits.  For example, the back end of a linear algebra package might benefit from value types like Float4 which map to stock vector types.  This is probably only worthwhile if the unboxing arrays can be packed with such values. More Daydreams A more finely-divided design for dynamic enforcement of value safety could feature separate marker interfaces for each invariant.  An empty marker interface Unsynchronizable could cause suitable exceptions for monitor instructions on objects in marked classes.  More radically, a Interchangeable marker interface could cause JVM primitives that are sensitive to object identity to raise exceptions; the strangest result would be that the acmp instruction would have to be specified as raising an exception. @ValueSafe public interface ValueType extends java.io.Serializable,         Unsynchronizable, Interchangeable { … public class Complex implements ValueType {     // inherits Serializable, Unsynchronizable, Interchangeable, @ValueSafe     … It seems possible that Integer and the other wrapper types could be retro-fitted as value-safe types.  This is a major change, since wrapper objects would be unsynchronizable and their references interchangeable.  It is likely that code which violates value-safety for wrapper types exists but is uncommon.  It is less plausible to retro-fit String, since the prominent operation String.intern is often used with value-unsafe code. We should also reconsider the distinction between boxed and unboxed values in code.  The design presented above obscures that distinction.  As another thought experiment, we could imagine making a first class distinction in the type system between boxed and unboxed representations.  Since only primitive types are named with a lower-case initial letter, we could define that the capitalized version of a value type name always refers to the boxed representation, while the initial lower-case variant always refers to boxed.  For example: complex pi = complex.valueOf(Math.PI, 0); Complex boxPi = pi;  // convert to boxed myList.add(boxPi); complex z = myList.get(0);  // unbox Such a convention could perhaps absorb the current difference between int and Integer, double and Double. It might also allow the programmer to express a helpful distinction among array types. As said above, array types are crucial to bulk data interfaces, but are limited in the JVM.  Extending arrays beyond the present limitations is worth thinking about; for example, the Maxine JVM implementation has a hybrid object/array type.  Something like this which can also accommodate value type components seems worthwhile.  On the other hand, does it make sense for value types to contain short arrays?  And why should random-access arrays be the end of our design process, when bulk data is often sequentially accessed, and it might make sense to have heterogeneous streams of data as the natural “jumbo” data structure.  These considerations must wait for another day and another note. More Work It seems to me that a good sequence for introducing such value types would be as follows: Add the value-safety restrictions to an experimental version of javac. Code some sample applications with value types, including Complex and DecimalValue. Create an experimental JVM which internally unboxes value types but does not require new bytecodes to do so.  Ensure the feasibility of the performance model for the sample applications. Add tuple-like bytecodes (with or without generic type reification) to a major revision of the JVM, and teach the Java compiler to switch in the new bytecodes without code changes. A staggered roll-out like this would decouple language changes from bytecode changes, which is always a convenient thing. A similar investigation should be applied (concurrently) to array types.  In this case, it seems to me that the starting point is in the JVM: Add an experimental unboxing array data structure to a production JVM, perhaps along the lines of Maxine hybrids.  No bytecode or language support is required at first; everything can be done with encapsulated unsafe operations and/or method handles. Create an experimental JVM which internally unboxes value types but does not require new bytecodes to do so.  Ensure the feasibility of the performance model for the sample applications. Add tuple-like bytecodes (with or without generic type reification) to a major revision of the JVM, and teach the Java compiler to switch in the new bytecodes without code changes. That’s enough musing me for now.  Back to work!

    Read the article

  • MP3 Decoding on Android

    - by Rob Szumlakowski
    Hi. We're implementing a program for Android phones that plays audio streamed from the internet. Here's approximately what we do: Download a custom encrypted format. Decrypt to get chunks of regular MP3 data. Decode MP3 data to raw PCM data in a memory buffer. Pipe the raw PCM data to an AudioTrack Our target devices so far are Droid and Nexus One. Everything works great on Nexus One, but the MP3 decode is too slow on Droid. The audio playback starts to skip if we put the Droid under load. We are not permitted to decode the MP3 data to SD card, but I know that's not our problem anyways. We didn't write our own MP3 decoder, but used MPADEC (http://sourceforge.net/projects/mpadec/). It's free and was easy to integrate with our program. We compile it with the NDK. After exhaustive analysis with various profiling tools, we're convinced that it's this decoder that is falling behind. Here's the options we're thinking about: Find another MP3 decoder that we can compile with the Android NDK. This MP3 decoder would have to be either optimized to run on mobile ARM devices or maybe use integer-only math or some other optimizations to increase performance. Since the built-in Android MediaPlayer service will take URLs, we might be able to implement a tiny HTTP server in our program and serve the MediaPlayer with the decrypted MP3s. That way we can take advantage of the built-in MP3 decoder. Get access to the built-in MP3 decoder through the NDK. I don't know if this is possible. Does anyone have any suggestions on what we can do to speed up our MP3 decoding? -- Rob Sz

    Read the article

  • .NET converting simple arrays to List Generics

    - by Manish Sinha
    This question might seem trivial and also stupid at the first glance, but it is much more than this. I have an array of any type T (T[]) and I want to convert it into a List generic (List<T>). Is there any other way apart from creating a Generic list, traversing the whole array and adding the element in the List? Present Situation: string[] strList = {'foo','bar','meh'}; List<string> listOfStr = new List<string>(); foreach(string s in strList) { listOfStr.Add(s); } My ideal situation: string[] strList = {'foo','bar','meh'}; List<string> listOfStr = strList.ToList<string>(); Or: string[] strList = {'foo','bar','meh'}; List<string> listOfStr = new List<string>(strList); I am suggesting the last 2 method names as I think compiler or CLR can perform some optimizations on the whole operations if It want inbuilt. P.S.: I am not talking about the Array or ArrayList Type

    Read the article

  • Optimal Serialization of Primitive Types

    - by Greg Dean
    We are beginning to roll out more and more WAN deployments of our product (.Net fat client w/ IIS hosted Remoting backend). Because of this we are trying to reduce the size of the data on the wire. We have overridden the default serialization by implementing ISerializable (similar to this), we are seeing anywhere from 12% to 50% gains. Most of our efforts focus on optimizing arrays of primitive types. I would like to know if anyone knows of any fancy way of serializing primitive types, beyond the obvious? For example today we serialize an array of ints as follows: [4-bytes (array length)][4-bytes][4-bytes] Can anyone do significantly better? The most obvious example of a significant improvement, for boolean arrays, is putting 8 bools in each byte, which we already do. Note: Saving 7 bits per bool may seem like a waste of time, but when you are dealing with large magnitudes of data (which we are), it adds up very fast. Note: We want to avoid general compression algorithms because of the latency associated with it. Remoting only supports buffered requests/responses(no chunked encoding). I realize there is a fine line between compression and optimal serialization, but our tests indicate we can afford very specific serialization optimizations at very little cost in latency. Whereas reprocessing the entire buffered response into new compressed buffer is too expensive.

    Read the article

  • P6 Architecture - Register renaming aside, does the limited user registers result in more ops spent

    - by mrjoltcola
    I'm studying JIT design with regard to dynamic languages VM implementation. I haven't done much Assembly since the 8086/8088 days, just a little here or there, so be nice if I'm out of sorts. As I understand it, the x86 (IA-32) architecture still has the same basic limited register set today that it always did, but the internal register count has grown tremendously, but these internal registers are not generally available and are used with register renaming to achieve parallel pipelining of code that otherwise could not be parallelizable. I understand this optimization pretty well, but my feeling is, while these optimizations help in overall throughput and for parallel algorithms, the limited register set we are still stuck with results in more register spilling overhead such that if x86 had double, or quadruple the registers available to us, there may be significantly less push/pop opcodes in a typical instruction stream? Or are there other processor optmizations that also optimize this away that I am unaware of? Basically if I've a unit of code that has 4 registers to work with for integer work, but my unit has a dozen variables, I've got potentially a push/pop for every 2 or so instructions. Any references to studies, or better yet, personal experiences?

    Read the article

  • Compiler optimization causing the performance to slow down

    - by aJ
    I have one strange problem. I have following piece of code: template<clss index, class policy> inline int CBase<index,policy>::func(const A& test_in, int* srcPtr ,int* dstPtr) { int width = test_in.width(); int height = test_in.height(); double d = 0.0; //here is the problem for(int y = 0; y < height; y++) { //Pointer initializations //multiplication involving y //ex: int z = someBigNumber*y + someOtherBigNumber; for(int x = 0; x < width; x++) { //multiplication involving x //ex: int z = someBigNumber*x + someOtherBigNumber; if(soemCondition) { // floating point calculations } *dstPtr++ = array[*srcPtr++]; } } } The inner loop gets executed nearly 200,000 times and the entire function takes 100 ms for completion. ( profiled using AQTimer) I found an unused variable double d = 0.0; outside the outer loop and removed the same. After this change, suddenly the method is taking 500ms for the same number of executions. ( 5 times slower). This behavior is reproducible in different machines with different processor types. (Core2, dualcore processors). I am using VC6 compiler with optimization level O2. Follwing are the other compiler options used : -MD -O2 -Z7 -GR -GX -G5 -X -GF -EHa I suspected compiler optimizations and removed the compiler optimization /O2. After that function became normal and it is taking 100ms as old code. Could anyone throw some light on this strange behavior? Why compiler optimization should slow down performance when I remove unused variable ? Note: The assembly code (before and after the change) looked same.

    Read the article

  • System.Dynamic bug?

    - by ControlFlow
    While I playing with the C# 4.0 dynamic, I found strange things happening with the code like this: using System.Dynamic; sealed class Foo : DynamicObject { public override bool TryInvoke( InvokeBinder binder, object[] args, out object result) { result = new object(); return true; } static void Main() { dynamic foo = new Foo(); var t1 = foo(0); var t2 = foo(0); var t3 = foo(0); var t4 = foo(0); var t5 = foo(0); } } Ok, it works but... take a look at IntelliTrace window: So every invokation (and other operations too on dynamic object) causes throwing and catching strange exceptions twice! I understand, that sometimes exceptions mechanism may be used for optimizations, for example first call to dynamic may be performed to some stub delegate, that simply throws exception - this may be like a signal to dynamic binder to resolve an correct member and re-point delegate. Next call to the same delegate will be performed without any checks. But... behavior of the code above looks very strange. Maybe throwing and catching exceptions twice per any operation on DynamicObject - is a bug?

    Read the article

  • Bubble sort algorithm implementations (Haskell vs. C)

    - by kingping
    Hello. I have written 2 implementation of bubble sort algorithm in C and Haskell. Haskell implementation: module Main where main = do contents <- readFile "./data" print "Data loaded. Sorting.." let newcontents = bubblesort contents writeFile "./data_new_ghc" newcontents print "Sorting done" bubblesort list = sort list [] False rev = reverse -- separated. To see rev2 = reverse -- who calls the routine sort (x1:x2:xs) acc _ | x1 > x2 = sort (x1:xs) (x2:acc) True sort (x1:xs) acc flag = sort xs (x1:acc) flag sort [] acc True = sort (rev acc) [] False sort _ acc _ = rev2 acc I've compared these two implementations having run both on file with size of 20 KiB. C implementation took about a second, Haskell — about 1 min 10 sec. I have also profiled the Haskell application: Compile for profiling: C:\Temp ghc -prof -auto-all -O --make Main Profile: C:\Temp Main.exe +RTS -p and got these results. This is a pseudocode of the algorithm: procedure bubbleSort( A : list of sortable items ) defined as: do swapped := false for each i in 0 to length(A) - 2 inclusive do: if A[i] > A[i+1] then swap( A[i], A[i+1] ) swapped := true end if end for while swapped end procedure I wonder if it's possible to make Haskell implementation work faster without changing the algorithm (there's are actually a few tricks to make it work faster, but neither implementations have these optimizations)

    Read the article

  • Transitive SQL query on same table

    - by MiKu
    Hey. consider d following table and data... in_timestamp | out_timestamp | name | in_id | out_id | in_server | out_server | status timestamp1 | timestamp2 | data1 |id1 | id2 | others-server1 | my-server1 | success timestamp2 | timestamp3 | data1 | id2 | id3 | my-server1 | my-server2 | success timestamp3 | timestamp4 | data1 | id3 | id4 | my-server2 | my-server3 | success timestamp4 | timestamp5 | data1 | id4 | id5 | my-server3 | others-server2 | success the above data represent log of a execution flow of some data across servers. e.g. some data has flowed from some 'outside-server1' to bunch of 'my-servers' and finally to destined 'others-server2'. Question : 1) I need to give this log in representable form to client where he doesn't need to know anything about the bunch of 'my-servers'. All i am supposed to give is timestamp of the data entered my infrastructure and when it left; drilling down to following info. in_timestamp (of 'others_server1' to 'my-server1') out_timestamp (of 'my-server3' to 'others-server2') name status I want to write sql for the same! Can someone help? NOTE : there might not be 3 'my-servers' all the time. It differs from situation to situation. e.g. there might be 4 'my-server' involved for, say, data2! 2) Are there any other alternatives to SQL? I mean stored procs/etc? 3) Optimizations? (The records are huge in number! As of now, it is around 5 million a day. And we are supposed to show records that are upto a week old.) In advance, THANKS FOR THE HELP! :)

    Read the article

  • F#: Tell me what I'm missing about using Async.Parallel

    - by JBristow
    ok, so I'm doing ProjectEuler Problem #14, and I'm fiddling around with optimizations in order to feel f# out. in the following code: let evenrule n = n / 2L let oddrule n = 3L * n + 1L let applyRule n = if n % 2L = 0L then evenrule n else oddrule n let runRules n = let rec loop a final = if a = 1L then final else loop (applyRule a) (final + 1L) n, loop (int64 n) 1L let testlist = seq {for i in 3 .. 2 .. 1000000 do yield i } let getAns sq = sq |> Seq.head let seqfil (a,acc) (b,curr) = if acc = curr then (a,acc) else if acc < curr then (b,curr) else (a,acc) let pmap f l = seq { for a in l do yield async {return f a} } |> Seq.map Async.RunSynchronously let pmap2 f l = seq { for a in l do yield async {return f a} } |> Async.Parallel |> Async.RunSynchronously let procseq f l = l |> f runRules |> Seq.reduce seqfil |> fst let timer = System.Diagnostics.Stopwatch() timer.Start() let ans1 = testlist |> procseq Seq.map // 837799 00:00:08.6251990 printfn "%A\t%A" ans1 timer.Elapsed timer.Reset() timer.Start() let ans2 = testlist |> procseq pmap printfn "%A\t%A" ans2 timer.Elapsed // 837799 00:00:12.3010250 timer.Reset() timer.Start() let ans3 = testlist |> procseq pmap2 printfn "%A\t%A" ans3 timer.Elapsed // 837799 00:00:58.2413990 timer.Reset() Why does the Async.Parallel code run REALLY slow in comparison to the straight up map? I know I shouldn't see that much of an effect, since I'm only on a dual core mac. Please note that I do NOT want help solving problem #14, I just want to know what's up with my parallel code.

    Read the article

  • Use of const double for intermediate results

    - by Arne
    Hi, I a writing a Simulation program and wondering if the use of const double is of any use when storing intermediate results. Consider this snippet: double DoSomeCalculation(const AcModel &model) { (...) const double V = model.GetVelocity(); const double m = model.GetMass(); const double cos_gamma = cos(model.GetFlightPathAngleRad()); (...) return m*V*cos_gamma*Chi_dot; } Note that the sample is there only to illustrate -- it might not make to much sense from the engineering side of things. The motivation of storing for example cos_gamma in a variable is that this cosine is used many time in other expressions covered by (...) and I feel that the code gets more readable when using cos_gamma rather than cos(model.GetFlightPathAngleRad()) in various expressions. Now the actual is question is this: since I expect the cosine to be the same througout the code section and I actually created the thing only as a placeholder and for convenience I tend to declare it const. Is there a etablished opinion on wether this is good or bad practive or whether it might bite me in the end? Does a compiler make any use of this additional information or am I actually hindering the compiler from performing useful optimizations? Arne

    Read the article

  • SQL: Speed Improvement - Cluttered union query

    - by vol7ron
    SELECT * FROM ( SELECT a.user_id, a.f_name, a.l_name, b.user_id, b.f_name, b.l_name FROM current_tbl a INNER JOIN import_tbl b ON ( a.user_id = b.user_id ) UNION SELECT a.user_id, a.f_name, a.l_name, b.user_id, b.f_name, b.l_name FROM current_tbl a INNER JOIN import_tbl b ON ( lower(a.f_name)=lower(b.f_name) AND lower(a.l_name)=lower(b.l_name) ) ) foo -- UNION -- SELECT a.user_id , a.f_name , a.l_name , '' , '' , '' FROM current_tbl a WHERE a.user_id NOT IN ( select user_id from( SELECT a.user_id, a.f_name, a.l_name, b.user_id, b.f_name, b.l_name FROM current_tbl a INNER JOIN import_tbl b ON ( a.user_id = b.user_id ) UNION SELECT a.user_id, a.f_name, a.l_name, b.user_id, b.f_name, b.l_name FROM current_tbl a INNER JOIN import_tbl b ON ( lower(a.f_name)=lower(b.f_name) AND lower(a.l_name)=lower(b.l_name) ) ) bar ) ORDER BY user_id Example of table population: current_tbl: ------------------------------- user_id | f_name | l_name ---------+----------+---------- A1 | Adam | Acorn A2 | Beth | Berry A3 | Calv | Chard | | import_tbl: ------------------------------- user_id | f_name | l_name ---------+----------+---------- A1 | Adam | Acorn A2 | Beth | Butcher <- last_name different | | Expected Output: ----------------------------------------------------------------------- user_id1 | f_name1 | l_name1 | user_id2 | f_name2 | l_name2 ----------+-----------+-----------+------------+-----------+----------- A1 | Adam | Acorn | A1 | Adam | Acorn A2 | Beth | Berry | A2 | Beth | Butcher A3 | Calv | Chard | | | Doing this method gets rid of conditions where the row would be: A2 | Beth | Berry | A2 | Beth | Butcher But it keeps the A3 row I hope this makes sense and I haven't overly simplified it. This is a continuation question from my other question. The succession of these improvements has dropped the query down from ~32000ms to where it's at now ~1200ms - quite an improvement. I supect I can optimize by using UNION ALL in the subquery and of course the usual index optimizations, but I'm looking for the best SQL optimization. FYI this particular case is for PostgreSQL.

    Read the article

  • [Doxygen] How to documenting global dependencies for functions?

    - by Thomas Matthews
    I've got some C code from a 3rd party vendor (for an embedded platform) that uses global variables (for speed & space optimizations). I'm documenting the code, converting to Doxygen format. How do I put a note in the function documentation that the function requires on global variables and functions? Doxygen has special commands for annotating parameters and return values as describe here: Doxygen Special Commands. I did not see any commands for global variables. Example C code: extern unsigned char data_buffer[]; //!< Global variable. /*! Returns the next available data byte. * \return Next data byte. */ unsigned char Get_Byte(void) { static unsigned int index = 0; return data_buffer[index++]; //!< Uses global variable. } In the above code, I would like to add Doxygen comments that the function depends on the global variable data_buffer.

    Read the article

  • Does this mimic perfectly a function template specialization?

    - by zeroes00
    Since the function template in the following code is a member of a class template, it can't be specialized without specializing the enclosing class. But if the compiler's full optimizations are on (assume Visual Studio 2010), will the if-else-statement in the following code get optimized out? And if it does, wouldn't it mean that for all practical purposes this IS a function template specialization without any performance cost? template<typename T> struct Holder { T data; template<int Number> void saveReciprocalOf(); }; template<typename T> template<int Number> void Holder<T>::saveReciprocalOf() { //Will this if-else-statement get completely optimized out if(Number == 0) data = (T)0; else data = (T)1 / Number; } //----------------------------------- void main() { Holder<float> holder; holder.saveReciprocalOf<2>(); cout << holder.data << endl; }

    Read the article

  • What is faster- Java or C# (Or good old C)?

    - by Rexsung
    I'm currently deciding on a platform to build a scientific computational product on, and am deciding on either C#, Java, or plain C with Intels compiler on Core2 Quad CPU's. It's mostly integer arithmetic. My benchmarks so far show Java and C are about on par with each other, and dotNET/C# trails by about 5%- however a number of my coworkers are claiming that dotNET with the right optimizations will beat both of these given enough time for the JIT to do its work. I always assume that the JIT would have done it's job within a few minutes of the app starting (Probably a few seconds in my case, as it's mostly tight loops), so I'm not sure whether to believe them Can anyone shed any light on the situation? Would dotNET beat Java? (Or am I best just sticking with C at this point?). The code is highly multithreaded and data sets are several terabytes in size. Haskell/erlang etc are not options in this case as there is a significant quantity of existing legacy C code that will be ported to the new system, and porting C to Java/C# is a lot simpler than to Haskell or Erlang. (Unless of course these provide a significant speedup). Edit: We are considering moving to C# or Java because they may, in theory, be faster. Every percent we can shave off our processing time saves us tens of thousands of dollars per year. At this point we are just trying to evaluate whether C, Java, or c# would be faster.

    Read the article

  • Writing a JavaScript zip code validation function

    - by mkoryak
    I would like to write a JavaScript function that validates a zip code, by checking if the zip code actually exists. Here is a list of all zip codes: http://www.census.gov/tiger/tms/gazetteer/zips.txt (I only care about the 2nd column) This is really a compression problem. I would like to do this for fun. OK, now that's out of the way, here is a list of optimizations over a straight hashtable that I can think of, feel free to add anything I have not thought of: Break zipcode into 2 parts, first 2 digits and last 3 digits. Make a giant if-else statement first checking the first 2 digits, then checking ranges within the last 3 digits. Or, covert the zips into hex, and see if I can do the same thing using smaller groups. Find out if within the range of all valid zip codes there are more valid zip codes vs invalid zip codes. Write the above code targeting the smaller group. Break up the hash into separate files, and load them via Ajax as user types in the zipcode. So perhaps break into 2 parts, first for first 2 digits, second for last 3. Lastly, I plan to generate the JavaScript files using another program, not by hand. Edit: performance matters here. I do want to use this, if it doesn't suck. Performance of the JavaScript code execution + download time. Edit 2: JavaScript only solutions please. I don't have access to the application server, plus, that would make this into a whole other problem =)

    Read the article

  • PHP templating challenge (optimizing front-end templates)

    - by Matt
    Hey all, I'm trying to do some templating optimizations and I'm wondering if it is possible to do something like this: function table_with_lowercase($data) { $out = '<table>'; for ($i=0; $i < 3; $i++) { $out .= '<tr><td>'; $out .= strtolower($data); $out .= '</td></tr>'; } $out .= "</table>"; return $out; } NOTE: You do not know what $data is when you run this function. Results in: <table> <tr><td><?php echo strtolower($data) ?></td></tr> <tr><td><?php echo strtolower($data) ?></td></tr> <tr><td><?php echo strtolower($data) ?></td></tr> </table> General Case: Anything that can be evaluated (compiled) will be. Any time there is an unknown variable, the variable and the functions enclosing it, will be output in a string format. Here's one more example: function capitalize($str) { return ucwords(strtolower($str)); } If $str is "HI ALL" then the output is: Hi All If $str is unknown then the output is: <?php echo ucwords(strtolower($str)); ?> In this case it would be easier to just call the function (ie. <?php echo capitalize($str) ?> ), but the example before would allow you to precompile your PHP to make it more efficient

    Read the article

  • PHP templating challenge (optimizing front-end templates)

    - by Matt
    Hey all, I'm trying to do some templating optimizations and I'm wondering if it is possible to do something like this: function table_with_lowercase($data) { $out = '<table>'; for ($i=0; $i < 3; $i++) { $out .= '<tr><td>'; $out .= strtolower($data); $out .= '</td></tr>'; } $out .= "</table>"; return $out; } NOTE: You do not know what $data is when you run this function. Results in: <table> <tr><td><?php echo strtolower($data) ?></td></tr> <tr><td><?php echo strtolower($data) ?></td></tr> <tr><td><?php echo strtolower($data) ?></td></tr> </table> General Case: Anything that can be evaluated (compiled) will be. Any time there is an unknown variable, the variable and the functions enclosing it, will be output in a string format. Here's one more example: function capitalize($str) { return ucwords(strtolower($str)); } If $str is "HI ALL" then the output is: Hi All If $str is unknown then the output is: <?php echo ucwords(strtolower($str)); ?> In this case it would be easier to just call the function (ie. <?php echo capitalize($str) ?> ), but the example before would allow you to precompile your PHP to make it more efficient

    Read the article

  • Need advice on comparing the performance of 2 equivalent linq to sql queries

    - by uvita
    I am working on tool to optimize linq to sql queries. Basically it intercepts the linq execution pipeline and makes some optimizations like for example removing a redundant join from a query. Of course, there is an overhead in the execution time before the query gets executed in the dbms, but then, the query should be processed faster. I don't want to use a sql profiler because I know that the generated query will be perform better in the dbms than the original one, I am looking for a correct way of measuring the global time between the creation of the query in linq and the end of its execution. Currently, I am using the Stopwatch class and my code looks something like this: var sw = new Stopwatch(); sw.Start(); const int amount = 100; for (var i = 0; i < amount; i++) { ExecuteNonOptimizedQuery(); } sw.Stop(); Console.Writeline("Executing the query {2} times took: {0}ms. On average, each query took: {1}ms", sw.ElapsedMilliseconds, sw.ElapsedMilliseconds / amount, amount); Basically the ExecutenNonOptimizedQuery() method creates a new DataContext, creates a query and then iterates over the results. I did this for both versions of the query, the normal one and the optimized one. I took the idea from this post from Frans Bouma. Is there any other approach/considerations I should take? Thanks in advance!

    Read the article

  • (x86) Assembler Optimization

    - by Pindatjuh
    I'm building a compiler/assembler/linker in Java for the x86-32 (IA32) processor targeting Windows. High-level concepts of a "language" (in essential a Java API for creating executables) are translated into opcodes, which then are wrapped and outputted to a file. The translation process has several phases, one is the translation between languages: the highest-level code is translated into the medium-level code which is then translated into the lowest-level code (probably more than 3 levels). My problem is the following; if I have higher-level code (X and Y) translated to lower-level code (x, y, U and V), then an example of such a translation is, in pseudo-code: x + U(f) // generated by X + V(f) + y // generated by Y (An easy example) where V is the opposite of U (compare with a stack push as U and a pop as V). This needs to be 'optimized' into: x + y (essentially removing the "useless" code) My idea was to use regular expressions. For the above case, it'll be a regular expression looking like this: x:(U(x)+V(x)):null, meaning for all x find U(x) followed by V(x) and replace by null. Imagine more complex regular expressions, for more complex optimizations. This should work on all levels. What do you suggest? What would be a good approach to optimize in these situations?

    Read the article

  • Optimization in Python - do's, don'ts and rules of thumb.

    - by JV
    Well I was reading this post and then I came across a code which was: jokes=range(1000000) domain=[(0,(len(jokes)*2)-i-1) for i in range(0,len(jokes)*2)] I thought wouldn't it be better to calculate the value of len(jokes) once outside the list comprehension? Well I tried it and timed three codes jv@Pioneer:~$ python -m timeit -s 'jokes=range(1000000);domain=[(0,(len(jokes)*2)-i-1) for i in range(0,len(jokes)*2)]' 10000000 loops, best of 3: 0.0352 usec per loop jv@Pioneer:~$ python -m timeit -s 'jokes=range(1000000);l=len(jokes);domain=[(0,(l*2)-i-1) for i in range(0,l*2)]' 10000000 loops, best of 3: 0.0343 usec per loop jv@Pioneer:~$ python -m timeit -s 'jokes=range(1000000);l=len(jokes)*2;domain=[(0,l-i-1) for i in range(0,l)]' 10000000 loops, best of 3: 0.0333 usec per loop Observing the marginal difference 2.55% between the first and the second made me think - is the first list comprehension domain=[(0,(len(jokes)*2)-i-1) for i in range(0,len(jokes)*2)] optimized internally by python? or is 2.55% a big enough optimization (given that the len(jokes)=1000000)? If this is - What are the other implicit/internal optimizations in Python ? What are the developer's rules of thumb for optimization in Python? Edit1: Since most of the answers are "don't optimize, do it later if its slow" and I got some tips and links from Triptych and Ali A for the do's. I will change the question a bit and request for don'ts. Can we have some experiences from people who faced the 'slowness', what was the problem and how it was corrected? Edit2: For those who haven't here is an interesting read Edit3: Incorrect usage of timeit in question please see dF's answer for correct usage and hence timings for the three codes.

    Read the article

  • sql-server performance optimization by removing print statements

    - by AG
    We're going through a round of sql-server stored procedure optimizations. The one recommendation we've found that clearly applies for us is 'SET NOCOUNT ON' at the top of each procedure. (Yes, I've seen the posts that point out issues with this depending on what client objects you run the stored procedures from but these are not issues for us.) So now I'm just trying to add in a bit of common sense. If the benefit of SET NOCOUNT ON is simply to reduce network traffic by some small amount every time, wouldn't it also make sense to turn off all the PRINT statements we have in the stored procedures that we only use for debugging? I can't see how it can hurt performance. OTOH, it's a bit of a hassle to implement due to the fact that some of the print statements are the only thing within else clauses, so you can't just always comment out the one line and be done. The change carries some amount of risk so I don't want to do it if it isn't going to actually help. But I don't see eliminating print statements mentioned anywhere in articles on optimization. Is that because it is so obvious no one bothers to mention it?

    Read the article

  • Collision detections and how efficient they are

    - by Shadow
    How exactly do you implement collision detection? What are the costs involved? Do different platforms(c/c++, java, cocoa/iphone, flash, directX) have different optimizations for calculating collisions. And lastly are there libraries available to do this for me, or some that I can just interpret for my platform of choice? As I understand it you would need to loop through the collision map and find the area in question and then compair the input thing(e.g. a sprite) to the type of pixel that is in the questioned area. I understand the very basic idea, but I don't understand the underlying implementation or even a higher level one for that matter. It would seem that this type of detection, or any for that matter, is very costly. Tile map? Bit array? How are these created from an image(I would guess looping and doing stuff)? The reason I ask this question is to get a better understanding of the efficiency behind the scenes and to understand exactly what is going on. Links, references, or examples would be very helpful. I know this question is a bit longwinded so any help or references would be very welcome. Thanks SO!

    Read the article

< Previous Page | 8 9 10 11 12 13 14 15 16 17 18 19  | Next Page >