C++ library profile Goal: Give performance improvement advice based on recognition of suboptimal usage patterns of the standard library. Method: Wrap the standard library code. Insert calls to an instrumentation library to record the internal state of various components at interesting entry/exit points to/from the standard library. Process trace, recognize suboptimal patterns, give advice. For details, see paper presented at CGO 2009. Strengths: Unintrusive solution. The application code does not require any modification. The advice is call context sensitive, thus capable of identifying precisely interesting dynamic performance behavior. The overhead model is pay-per-view. When you turn off a diagnostic class at compile time, its overhead disappears. Drawbacks: You must recompile the application code with custom options. You must run the application on representative input. The advice is input dependent. The execution time will increase, in some cases by factors. This is the anticipated common workflow for program foo.cc: $ cat foo.cc #include <vector> int main() { vector<int> v; for (int k = 0; k < 1024; ++k) v.insert(v.begin(), k); } $ g++ -D_GLIBCXX_PROFILE foo.cc $ ./a.out $ cat libstdcxx-profile.txt vector-to-list: improvement = 5: call stack = 0x804842c ... : advice = change std::vector to std::list vector-size: improvement = 3: call stack = 0x804842c ... : advice = change initial container size from 0 to 1024 Anatomy of a warning: Warning id. This is a short descriptive string for the class that this warning belongs to. E.g., "vector-to-list". Estimated improvement. This is an approximation of the benefit expected from implementing the change suggested by the warning. It is given on a log10 scale. Negative values mean that the alternative would actually do worse than the current choice. In the example above, 5 comes from the fact that the overhead of inserting at the beginning of a vector vs. a list is around 1024 * 1024 / 2, which is around 10e5. The improvement from setting the initial size to 1024 is in the range of 10e3, since the overhead of dynamic resizing is linear in this case. Call stack. Currently, the addresses are printed without symbol name or code location attribution. Users are expected to postprocess the output using, for instance, addr2line. The warning message. For some warnings, this is static text, e.g., "change vector to list". For other warnings, such as the one above, the message contains numeric advice, e.g., the suggested initial size of the vector. Three files are generated. libstdcxx-profile.txt contains human readable advice. libstdcxx-profile.raw contains implementation specific data about each diagnostic. Their format is not documented. They are sufficient to generate all the advice given in libstdcxx-profile.txt. The advantage of keeping this raw format is that traces from multiple executions can be aggregated simply by concatenating the raw traces. We intend to offer an external utility program that can issue advice from a trace. libstdcxx-profile.conf.out lists the actual diagnostic parameters used. To alter parameters, edit this file and rename it to libstdcxx-profile.conf. Advice is given regardless whether the transformation is valid. For instance, we advise changing a map to an unordered_map even if the application semantics require that data be ordered. We believe such warnings can help users understand the performance behavior of their application better, which can lead to changes at a higher abstraction level. Compile time switches and environment variables (see also file profiler.h). Unless specified otherwise, they can be set at compile time using -D_<name> or by setting variable <name> in the environment where the program is run, before starting execution. _GLIBCXX_PROFILE_NO_<diagnostic>: disable specific diagnostics. See section Diagnostics for possible values. (Environment variables not supported.) _GLIBCXX_PROFILE_TRACE_PATH_ROOT: set an alternative root path for the output files. _GLIBCXX_PROFILE_MAX_WARN_COUNT: set it to the maximum number of warnings desired. The default value is 10. _GLIBCXX_PROFILE_MAX_STACK_DEPTH: if set to 0, the advice will be collected and reported for the program as a whole, and not for each call context. This could also be used in continuous regression tests, where you just need to know whether there is a regression or not. The default value is 32. _GLIBCXX_PROFILE_MEM_PER_DIAGNOSTIC: set a limit on how much memory to use for the accounting tables for each diagnostic type. When this limit is reached, new events are ignored until the memory usage decreases under the limit. Generally, this means that newly created containers will not be instrumented until some live containers are deleted. The default is 128 MB. _GLIBCXX_PROFILE_NO_THREADS: Make the library not use threads. If thread local storage (TLS) is not available, you will get a preprocessor error asking you to set -D_GLIBCXX_PROFILE_NO_THREADS if your program is single-threaded. Multithreaded execution without TLS is not supported. (Environment variable not supported.) _GLIBCXX_HAVE_EXECINFO_H: This name should be defined automatically at library configuration time. If your library was configured without execinfo.h, but you have it in your include path, you can define it explicitly. Without it, advice is collected for the program as a whole, and not for each call context. (Environment variable not supported.) Code Location Use libstdc++-v3/include/std/* Preprocessor code to redirect to profile extension headers. libstdc++-v3/include/profile/* Profile extension public headers (map, vector, ...). libstdc++-v3/include/profile/impl/* Profile extension internals. Implementation files are only included from impl/profiler.h, which is the only file included from the public headers. In order to get our instrumented library version included instead of the release one, we use the same wrapper model as the debug mode. We subclass entities from the release version. Wherever _GLIBCXX_PROFILE is defined, the release namespace is std::__norm, whereas the profile namespace is std::__profile. Using plain std translates into std::__profile. Whenever possible, we try to wrap at the public interface level, e.g., in unordered_set rather than in hashtable, in order not to depend on implementation. Mixing object files built with and without the profile mode must not affect the program execution. However, there are no guarantees to the accuracy of diagnostics when using even a single object not built with -D_GLIBCXX_PROFILE. Currently, mixing the profile mode with debug and parallel extensions is not allowed. Mixing them at compile time will result in preprocessor errors. Mixing them at link time is undefined. Instead of instrumenting every public entry and exit point, we chose to add instrumentation on demand, as needed by individual diagnostics. The main reason is that some diagnostics require us to extract bits of internal state that are particular only to that diagnostic. We plan to formalize this later, after we learn more about the requirements of several diagnostics. All the instrumentation points can be switched on and off using -D[_NO]_GLIBCXX_PROFILE_<diagnostic> options. With all the instrumentation calls off, there should be negligible overhead over the release version. This property is needed to support diagnostics based on timing of internal operations. For such diagnostics, we anticipate turning most of the instrumentation off in order to prevent profiling overhead from polluting time measurements, and thus diagnostics. All the instrumentation on/off compile time switches live in include/profile/profiler.h. For practical reasons, the instrumentation library processes the trace partially rather than dumping it to disk in raw form. Each event is processed when it occurs. It is usually attached a cost and it is aggregated into the database of a specific diagnostic class. The cost model is based largely on the standard performance guarantees, but in some cases we use knowledge about GCC's standard library implementation. Information is indexed by (1) call stack and (2) instance id or address to be able to understand and summarize precise creation-use-destruction dynamic chains. Although the analysis is sensitive to dynamic instances, the reports are only sensitive to call context. Whenever a dynamic instance is destroyed, we accumulate its effect to the corresponding entry for the call stack of its constructor location. For details, see paper presented at CGO 2009. Final analysis takes place offline, and it is based entirely on the generated trace and debugging info in the application binary. See section Diagnostics for a list of analysis types that we plan to support. The input to the analysis is a table indexed by profile type and call stack. The data type for each entry depends on the profile type. While it is likely that cost models become complex as we get into more sophisticated analysis, we will try to follow a simple set of rules at the beginning. Relative benefit estimation: The idea is to estimate or measure the cost of all operations in the original scenario versus the scenario we advise to switch to. For instance, when advising to change a vector to a list, an occurrence of the insert method will generally count as a benefit. Its magnitude depends on (1) the number of elements that get shifted and (2) whether it triggers a reallocation. Synthetic measurements: We will measure the relative difference between similar operations on different containers. We plan to write a battery of small tests that compare the times of the executions of similar methods on different containers. The idea is to run these tests on the target machine. If this training phase is very quick, we may decide to perform it at library initialization time. The results can be cached on disk and reused across runs. Timers: We plan to use timers for operations of larger granularity, such as sort. For instance, we can switch between different sort methods on the fly and report the one that performs best for each call context. Show stoppers: We may decide that the presence of an operation nullifies the advice. For instance, when considering switching from set to unordered_set, if we detect use of operator ++, we will simply not issue the advice, since this could signal that the use care require a sorted container. There are two types of reports. First, if we recognize a pattern for which we have a substitute that is likely to give better performance, we print the advice and estimated performance gain. The advice is usually associated to a code position and possibly a call stack. Second, we report performance characteristics for which we do not have a clear solution for improvement. For instance, we can point to the user the top 10 multimap locations which have the worst data locality in actual traversals. Although this does not offer a solution, it helps the user focus on the key problems and ignore the uninteresting ones. First, we want to make sure we preserve the behavior of the release mode. You can just type "make check-profile", which builds and runs the whole test suite in profile mode. Second, we want to test the correctness of each diagnostic. We created a profile directory in the test suite. Each diagnostic must come with at least two tests, one for false positives and one for false negatives. Many large projects use their own data structures instead of the ones in the standard library. If these data structures are similar in functionality to the standard library, they can be instrumented with the same hooks that are used to instrument the standard library. The instrumentation API is exposed in file profiler.h (look for "Instrumentation hooks"). Currently, the cost model uses formulas with predefined relative weights for alternative containers or container implementations. For instance, iterating through a vector is X times faster than iterating through a list. (Under development.) We are working on customizing this to a particular machine by providing an automated way to compute the actual relative weights for operations on the given machine. (Under development.) We plan to provide a performance parameter database format that can be filled in either by hand or by an automated training mechanism. The analysis module will then use this database instead of the built in. generic parameters. Accurate stack traces are needed during profiling since we group events by call context and dynamic instance. Without accurate traces, diagnostics may be hard to interpret. For instance, when giving advice to the user it is imperative to reference application code, not library code. Currently we are using the libc backtrace routine to get stack traces. _GLIBCXX_PROFILE_STACK_DEPTH can be set to 0 if you are willing to give up call context information, or to a small positive value to reduce run time overhead. The profiling and analysis phases use only instruction addresses. An external utility such as addr2line is needed to postprocess the result. We do not plan to add symbolization support in the profile extension. This would require access to symbol tables, debug information tables, external programs or libraries and other system dependent information. Our current model is simplistic, but precise. We cannot afford to approximate because some of our diagnostics require precise matching of operations to container instance and call context. During profiling, we keep a single information table per diagnostic. There is a single lock per information table. As much as we would like to avoid uses of libstdc++ within our instrumentation library, containers such as unordered_map are very appealing. We plan to use them as long as they are named properly to avoid ambiguity. User applications/libraries can provide malloc hooks. When the implementation of the malloc hooks uses stdlibc++, there can be an infinite cycle between the profile mode instrumentation and the malloc hook code. We protect against reentrance to the profile mode instrumentation code, which should avoid this problem in most cases. The protection mechanism is thread safe and exception safe. This mechanism does not prevent reentrance to the malloc hook itself, which could still result in deadlock, if, for instance, the malloc hook uses non-recursive locks. XXX: A definitive solution to this problem would be for the profile extension to use a custom allocator internally, and perhaps not to use libstdc++. The profiling library state is initialized at the first call to a profiling method. This allows us to record the construction of all global objects. However, we cannot do the same at destruction time. The trace is written by a function registered by atexit, thus invoked by exit. The profile mode headers are included with -D_GLIBCXX_PROFILE through preprocessor directives in include/std/*. Instrumented implementations are provided in include/profile/*. All instrumentation hooks are macros defined in include/profile/profiler.h. All the implementation of the instrumentation hooks is in include/profile/impl/*. Although all the code gets included, thus is publicly visible, only a small number of functions are called from outside this directory. All calls to hook implementations must be done through macros defined in profiler.h. The macro must ensure (1) that the call is guarded against reentrance and (2) that the call can be turned off at compile time using a -D_GLIBCXX_PROFILE_... compiler option. Let's say the diagnostic name is "magic". If you need to instrument a header not already under include/profile/*, first edit the corresponding header under include/std/ and add a preprocessor directive such as the one in include/std/vector: #ifdef _GLIBCXX_PROFILE # include <profile/vector> #endif If the file you need to instrument is not yet under include/profile/, make a copy of the one in include/debug, or the main implementation. You'll need to include the main implementation and inherit the classes you want to instrument. Then define the methods you want to instrument, define the instrumentation hooks and add calls to them. Look at include/profile/vector for an example. Add macros for the instrumentation hooks in include/profile/impl/profiler.h. Hook names must start with __profcxx_. Make sure they transform in no code with -D_NO_GLBICXX_PROFILE_MAGIC. Make sure all calls to any method in namespace __gnu_profile is protected against reentrance using macro _GLIBCXX_PROFILE_REENTRANCE_GUARD. All names of methods in namespace __gnu_profile called from profiler.h must start with __trace_magic_. Add the implementation of the diagnostic. Create new file include/profile/impl/profiler_magic.h. Define class __magic_info: public __object_info_base. This is the representation of a line in the object table. The __merge method is used to aggregate information across all dynamic instances created at the same call context. The __magnitude must return the estimation of the benefit as a number of small operations, e.g., number of words copied. The __write method is used to produce the raw trace. The __advice method is used to produce the advice string. Define class __magic_stack_info: public __magic_info. This defines the content of a line in the stack table. Define class __trace_magic: public __trace_base<__magic_info, __magic_stack_info>. It defines the content of the trace associated with this diagnostic. Add initialization and reporting calls in include/profile/impl/profiler_trace.h. Use __trace_vector_to_list as an example. Add documentation in file doc/xml/manual/profile_mode.xml. The table below presents all the diagnostics we intend to implement. Each diagnostic has a corresponding compile time switch -D_GLIBCXX_PROFILE_<diagnostic>. Groups of related diagnostics can be turned on with a single switch. For instance, -D_GLIBCXX_PROFILE_LOCALITY is equivalent to -D_GLIBCXX_PROFILE_SOFTWARE_PREFETCH -D_GLIBCXX_PROFILE_RBTREE_LOCALITY. The benefit, cost, expected frequency and accuracy of each diagnostic was given a grade from 1 to 10, where 10 is highest. A high benefit means that, if the diagnostic is accurate, the expected performance improvement is high. A high cost means that turning this diagnostic on leads to high slowdown. A high frequency means that we expect this to occur relatively often. A high accuracy means that the diagnostic is unlikely to be wrong. These grades are not perfect. They are just meant to guide users with specific needs or time budgets. Group Flag Benefit Cost Freq. Implemented CONTAINERS HASHTABLE_TOO_SMALL 10 1 10 yes HASHTABLE_TOO_LARGE 5 1 10 yes INEFFICIENT_HASH 7 3 10 yes VECTOR_TOO_SMALL 8 1 10 yes VECTOR_TOO_LARGE 5 1 10 yes VECTOR_TO_HASHTABLE 7 7 10 no HASHTABLE_TO_VECTOR 7 7 10 no VECTOR_TO_LIST 8 5 10 yes LIST_TO_VECTOR 10 5 10 no ORDERED_TO_UNORDERED 10 5 10 only map/unordered_map ALGORITHMS SORT 7 8 7 no LOCALITY SOFTWARE_PREFETCH 8 8 5 no RBTREE_LOCALITY 4 8 5 no FALSE_SHARING 8 10 10 no Switch: _GLIBCXX_PROFILE_<diagnostic>. Goal: What problem will it diagnose? Fundamentals:. What is the fundamental reason why this is a problem Sample runtime reduction: Percentage reduction in execution time. When reduction is more than a constant factor, describe the reduction rate formula. Recommendation: What would the advise look like? To instrument: What stdlibc++ components need to be instrumented? Analysis: How do we decide when to issue the advice? Cost model: How do we measure benefits? Math goes here. Example: program code ... advice sample Switch: _GLIBCXX_PROFILE_CONTAINERS. Switch: _GLIBCXX_PROFILE_HASHTABLE_TOO_SMALL. Goal: Detect hashtables with many rehash operations, small construction size and large destruction size. Fundamentals: Rehash is very expensive. Read content, follow chains within bucket, evaluate hash function, place at new location in different order. Sample runtime reduction: 36%. Code similar to example below. Recommendation: Set initial size to N at construction site S. To instrument: unordered_set, unordered_map constructor, destructor, rehash. Analysis: For each dynamic instance of unordered_[multi]set|map, record initial size and call context of the constructor. Record size increase, if any, after each relevant operation such as insert. Record the estimated rehash cost. Cost model: Number of individual rehash operations * cost per rehash. Example: 1 unordered_set<int> us; 2 for (int k = 0; k < 1000000; ++k) { 3 us.insert(k); 4 } foo.cc:1: advice: Changing initial unordered_set size from 10 to 1000000 saves 1025530 rehash operations. Switch: _GLIBCXX_PROFILE_HASHTABLE_TOO_LARGE. Goal: Detect hashtables which are never filled up because fewer elements than reserved are ever inserted. Fundamentals: Save memory, which is good in itself and may also improve memory reference performance through fewer cache and TLB misses. Sample runtime reduction: unknown. Recommendation: Set initial size to N at construction site S. To instrument: unordered_set, unordered_map constructor, destructor, rehash. Analysis: For each dynamic instance of unordered_[multi]set|map, record initial size and call context of the constructor, and correlate it with its size at destruction time. Cost model: Number of iteration operations + memory saved. Example: 1 vector<unordered_set<int>> v(100000, unordered_set<int>(100)) ; 2 for (int k = 0; k < 100000; ++k) { 3 for (int j = 0; j < 10; ++j) { 4 v[k].insert(k + j); 5 } 6 } foo.cc:1: advice: Changing initial unordered_set size from 100 to 10 saves N bytes of memory and M iteration steps. Switch: _GLIBCXX_PROFILE_INEFFICIENT_HASH. Goal: Detect hashtables with polarized distribution. Fundamentals: A non-uniform distribution may lead to long chains, thus possibly increasing complexity by a factor up to the number of elements. Sample runtime reduction: factor up to container size. Recommendation: Change hash function for container built at site S. Distribution score = N. Access score = S. Longest chain = C, in bucket B. To instrument: unordered_set, unordered_map constructor, destructor, [], insert, iterator. Analysis: Count the exact number of link traversals. Cost model: Total number of links traversed. Example: class dumb_hash { public: size_t operator() (int i) const { return 0; } }; ... unordered_set<int, dumb_hash> hs; ... for (int i = 0; i < COUNT; ++i) { hs.find(i); } Switch: _GLIBCXX_PROFILE_VECTOR_TOO_SMALL. Goal:Detect vectors with many resize operations, small construction size and large destruction size.. Fundamentals:Resizing can be expensive. Copying large amounts of data takes time. Resizing many small vectors may have allocation overhead and affect locality. Sample runtime reduction:%. Recommendation: Set initial size to N at construction site S. To instrument:vector. Analysis: For each dynamic instance of vector, record initial size and call context of the constructor. Record size increase, if any, after each relevant operation such as push_back. Record the estimated resize cost. Cost model: Total number of words copied * time to copy a word. Example: 1 vector<int> v; 2 for (int k = 0; k < 1000000; ++k) { 3 v.push_back(k); 4 } foo.cc:1: advice: Changing initial vector size from 10 to 1000000 saves copying 4000000 bytes and 20 memory allocations and deallocations. Switch: _GLIBCXX_PROFILE_VECTOR_TOO_LARGE Goal:Detect vectors which are never filled up because fewer elements than reserved are ever inserted. Fundamentals:Save memory, which is good in itself and may also improve memory reference performance through fewer cache and TLB misses. Sample runtime reduction:%. Recommendation: Set initial size to N at construction site S. To instrument:vector. Analysis: For each dynamic instance of vector, record initial size and call context of the constructor, and correlate it with its size at destruction time. Cost model: Total amount of memory saved. Example: 1 vector<vector<int>> v(100000, vector<int>(100)) ; 2 for (int k = 0; k < 100000; ++k) { 3 for (int j = 0; j < 10; ++j) { 4 v[k].insert(k + j); 5 } 6 } foo.cc:1: advice: Changing initial vector size from 100 to 10 saves N bytes of memory and may reduce the number of cache and TLB misses. Switch: _GLIBCXX_PROFILE_VECTOR_TO_HASHTABLE. Goal: Detect uses of vector that can be substituted with unordered_set to reduce execution time. Fundamentals: Linear search in a vector is very expensive, whereas searching in a hashtable is very quick. Sample runtime reduction:factor up to container size. Recommendation:Replace vector with unordered_set at site S. To instrument:vector operations and access methods. Analysis: For each dynamic instance of vector, record call context of the constructor. Issue the advice only if the only methods called on this vector are push_back, insert and find. Cost model: Cost(vector::push_back) + cost(vector::insert) + cost(find, vector) - cost(unordered_set::insert) + cost(unordered_set::find). Example: 1 vector<int> v; ... 2 for (int i = 0; i < 1000; ++i) { 3 find(v.begin(), v.end(), i); 4 } foo.cc:1: advice: Changing "vector" to "unordered_set" will save about 500,000 comparisons. Switch: _GLIBCXX_PROFILE_HASHTABLE_TO_VECTOR. Goal: Detect uses of unordered_set that can be substituted with vector to reduce execution time. Fundamentals: Hashtable iterator is slower than vector iterator. Sample runtime reduction:95%. Recommendation:Replace unordered_set with vector at site S. To instrument:unordered_set operations and access methods. Analysis: For each dynamic instance of unordered_set, record call context of the constructor. Issue the advice only if the number of find, insert and [] operations on this unordered_set are small relative to the number of elements, and methods begin or end are invoked (suggesting iteration). Cost model: Number of . Example: 1 unordered_set<int> us; ... 2 int s = 0; 3 for (unordered_set<int>::iterator it = us.begin(); it != us.end(); ++it) { 4 s += *it; 5 } foo.cc:1: advice: Changing "unordered_set" to "vector" will save about N indirections and may achieve better data locality. Switch: _GLIBCXX_PROFILE_VECTOR_TO_LIST. Goal: Detect cases where vector could be substituted with list for better performance. Fundamentals: Inserting in the middle of a vector is expensive compared to inserting in a list. Sample runtime reduction:factor up to container size. Recommendation:Replace vector with list at site S. To instrument:vector operations and access methods. Analysis: For each dynamic instance of vector, record the call context of the constructor. Record the overhead of each insert operation based on current size and insert position. Report instance with high insertion overhead. Cost model: (Sum(cost(vector::method)) - Sum(cost(list::method)), for method in [push_back, insert, erase]) + (Cost(iterate vector) - Cost(iterate list)) Example: 1 vector<int> v; 2 for (int i = 0; i < 10000; ++i) { 3 v.insert(v.begin(), i); 4 } foo.cc:1: advice: Changing "vector" to "list" will save about 5,000,000 operations. Switch: _GLIBCXX_PROFILE_LIST_TO_VECTOR. Goal: Detect cases where list could be substituted with vector for better performance. Fundamentals: Iterating through a vector is faster than through a list. Sample runtime reduction:64%. Recommendation:Replace list with vector at site S. To instrument:vector operations and access methods. Analysis: Issue the advice if there are no insert operations. Cost model: (Sum(cost(vector::method)) - Sum(cost(list::method)), for method in [push_back, insert, erase]) + (Cost(iterate vector) - Cost(iterate list)) Example: 1 list<int> l; ... 2 int sum = 0; 3 for (list<int>::iterator it = l.begin(); it != l.end(); ++it) { 4 sum += *it; 5 } foo.cc:1: advice: Changing "list" to "vector" will save about 1000000 indirect memory references. Switch: _GLIBCXX_PROFILE_LIST_TO_SLIST. Goal: Detect cases where list could be substituted with forward_list for better performance. Fundamentals: The memory footprint of a forward_list is smaller than that of a list. This has beneficial effects on memory subsystem, e.g., fewer cache misses. Sample runtime reduction:40%. Note that the reduction is only noticeable if the size of the forward_list node is in fact larger than that of the list node. For memory allocators with size classes, you will only notice an effect when the two node sizes belong to different allocator size classes. Recommendation:Replace list with forward_list at site S. To instrument:list operations and iteration methods. Analysis: Issue the advice if there are no backwards traversals or insertion before a given node. Cost model: Always true. Example: 1 list<int> l; ... 2 int sum = 0; 3 for (list<int>::iterator it = l.begin(); it != l.end(); ++it) { 4 sum += *it; 5 } foo.cc:1: advice: Change "list" to "forward_list". Switch: _GLIBCXX_PROFILE_ORDERED_TO_UNORDERED. Goal: Detect cases where ordered associative containers can be replaced with unordered ones. Fundamentals: Insert and search are quicker in a hashtable than in a red-black tree. Sample runtime reduction:52%. Recommendation: Replace set with unordered_set at site S. To instrument: set, multiset, map, multimap methods. Analysis: Issue the advice only if we are not using operator ++ on any iterator on a particular [multi]set|map. Cost model: (Sum(cost(hashtable::method)) - Sum(cost(rbtree::method)), for method in [insert, erase, find]) + (Cost(iterate hashtable) - Cost(iterate rbtree)) Example: 1 set<int> s; 2 for (int i = 0; i < 100000; ++i) { 3 s.insert(i); 4 } 5 int sum = 0; 6 for (int i = 0; i < 100000; ++i) { 7 sum += *s.find(i); 8 } Switch: _GLIBCXX_PROFILE_ALGORITHMS. Switch: _GLIBCXX_PROFILE_SORT. Goal: Give measure of sort algorithm performance based on actual input. For instance, advise Radix Sort over Quick Sort for a particular call context. Fundamentals: See papers: A framework for adaptive algorithm selection in STAPL and Optimizing Sorting with Machine Learning Algorithms. Sample runtime reduction:60%. Recommendation: Change sort algorithm at site S from X Sort to Y Sort. To instrument: sort algorithm. Analysis: Issue the advice if the cost model tells us that another sort algorithm would do better on this input. Requires us to know what algorithm we are using in our sort implementation in release mode. Cost model: Runtime(algo) for algo in [radix, quick, merge, ...] Example: Switch: _GLIBCXX_PROFILE_LOCALITY. Switch: _GLIBCXX_PROFILE_SOFTWARE_PREFETCH. Goal: Discover sequences of indirect memory accesses that are not regular, thus cannot be predicted by hardware prefetchers. Fundamentals: Indirect references are hard to predict and are very expensive when they miss in caches. Sample runtime reduction:25%. Recommendation: Insert prefetch instruction. To instrument: Vector iterator and access operator []. Analysis: First, get cache line size and page size from system. Then record iterator dereference sequences for which the value is a pointer. For each sequence within a container, issue a warning if successive pointer addresses are not within cache lines and do not form a linear pattern (otherwise they may be prefetched by hardware). If they also step across page boundaries, make the warning stronger. The same analysis applies to containers other than vector. However, we cannot give the same advice for linked structures, such as list, as there is no random access to the n-th element. The user may still be able to benefit from this information, for instance by employing frays (user level light weight threads) to hide the latency of chasing pointers. This analysis is a little oversimplified. A better cost model could be created by understanding the capability of the hardware prefetcher. This model could be trained automatically by running a set of synthetic cases. Cost model: Total distance between pointer values of successive elements in vectors of pointers. Example: 1 int zero = 0; 2 vector<int*> v(10000000, &zero); 3 for (int k = 0; k < 10000000; ++k) { 4 v[random() % 10000000] = new int(k); 5 } 6 for (int j = 0; j < 10000000; ++j) { 7 count += (*v[j] == 0 ? 0 : 1); 8 } foo.cc:7: advice: Insert prefetch instruction. Switch: _GLIBCXX_PROFILE_RBTREE_LOCALITY. Goal: Give measure of locality of objects stored in linked structures (lists, red-black trees and hashtables) with respect to their actual traversal patterns. Fundamentals:Allocation can be tuned to a specific traversal pattern, to result in better data locality. See paper: Custom Memory Allocation for Free. Sample runtime reduction:30%. Recommendation: High scatter score N for container built at site S. Consider changing allocation sequence or choosing a structure conscious allocator. To instrument: Methods of all containers using linked structures. Analysis: First, get cache line size and page size from system. Then record the number of successive elements that are on different line or page, for each traversal method such as find. Give advice only if the ratio between this number and the number of total node hops is above a threshold. Cost model: Sum(same_cache_line(this,previous)) Example: 1 set<int> s; 2 for (int i = 0; i < 10000000; ++i) { 3 s.insert(i); 4 } 5 set<int> s1, s2; 6 for (int i = 0; i < 10000000; ++i) { 7 s1.insert(i); 8 s2.insert(i); 9 } ... // Fast, better locality. 10 for (set<int>::iterator it = s.begin(); it != s.end(); ++it) { 11 sum += *it; 12 } // Slow, elements are further apart. 13 for (set<int>::iterator it = s1.begin(); it != s1.end(); ++it) { 14 sum += *it; 15 } foo.cc:5: advice: High scatter score NNN for set built here. Consider changing the allocation sequence or switching to a structure conscious allocator. The diagnostics in this group are not meant to be implemented short term. They require compiler support to know when container elements are written to. Instrumentation can only tell us when elements are referenced. Switch: _GLIBCXX_PROFILE_MULTITHREADED. Switch: _GLIBCXX_PROFILE_DDTEST. Goal: Detect container elements that are referenced from multiple threads in the parallel region or across parallel regions. Fundamentals: Sharing data between threads requires communication and perhaps locking, which may be expensive. Sample runtime reduction:?%. Recommendation: Change data distribution or parallel algorithm. To instrument: Container access methods and iterators. Analysis: Keep a shadow for each container. Record iterator dereferences and container member accesses. Issue advice for elements referenced by multiple threads. See paper: The LRPD test: speculative run-time parallelization of loops with privatization and reduction parallelization. Cost model: Number of accesses to elements referenced from multiple threads Example: Switch: _GLIBCXX_PROFILE_FALSE_SHARING. Goal: Detect elements in the same container which share a cache line, are written by at least one thread, and accessed by different threads. Fundamentals: Under these assumptions, cache protocols require communication to invalidate lines, which may be expensive. Sample runtime reduction:68%. Recommendation: Reorganize container or use padding to avoid false sharing. To instrument: Container access methods and iterators. Analysis: First, get the cache line size. For each shared container, record all the associated iterator dereferences and member access methods with the thread id. Compare the address lists across threads to detect references in two different threads to the same cache line. Issue a warning only if the ratio to total references is significant. Do the same for iterator dereference values if they are pointers. Cost model: Number of accesses to same cache line from different threads. Example: 1 vector<int> v(2, 0); 2 #pragma omp parallel for shared(v, SIZE) schedule(static, 1) 3 for (i = 0; i < SIZE; ++i) { 4 v[i % 2] += i; 5 } OMP_NUM_THREADS=2 ./a.out foo.cc:1: advice: Change container structure or padding to avoid false sharing in multithreaded access at foo.cc:4. Detected N shared cache lines. Switch: _GLIBCXX_PROFILE_STATISTICS. In some cases the cost model may not tell us anything because the costs appear to offset the benefits. Consider the choice between a vector and a list. When there are both inserts and iteration, an automatic advice may not be issued. However, the programmer may still be able to make use of this information in a different way. This diagnostic will not issue any advice, but it will print statistics for each container construction site. The statistics will contain the cost of each operation actually performed on the container. Lixia Liu Silvius Rus 2009 Proceedings of the 2009 International Symposium on Code Generation and Optimization