URL
https://opencores.org/ocsvn/openrisc/openrisc/trunk
Subversion Repositories openrisc
[/] [openrisc/] [trunk/] [gnu-dev/] [or1k-gcc/] [boehm-gc/] [doc/] [scale.html] - Rev 760
Go to most recent revision | Compare with Previous | Blame | View Log
<HTML> <HEAD> <TITLE>Garbage collector scalability</TITLE> </HEAD> <BODY> <H1>Garbage collector scalability</h1> In its default configuration, the Boehm-Demers-Weiser garbage collector is not thread-safe. It can be made thread-safe for a number of environments by building the collector with the appropriate <TT>-D</tt><I>XXX</i><TT>-THREADS</tt> compilation flag. This has primarily two effects: <OL> <LI> It causes the garbage collector to stop all other threads when it needs to see a consistent memory state. <LI> It causes the collector to acquire a lock around essentially all allocation and garbage collection activity. </ol> Since a single lock is used for all allocation-related activity, only one thread can be allocating or collecting at one point. This inherently limits performance of multi-threaded applications on multiprocessors. <P> On most platforms, the allocator/collector lock is implemented as a spin lock with exponential back-off. Longer wait times are implemented by yielding and/or sleeping. If a collection is in progress, the pure spinning stage is skipped. This has the advantage that uncontested and thus most uniprocessor lock acquisitions are very cheap. It has the disadvantage that the application may sleep for small periods of time even when there is work to be done. And threads may be unnecessarily woken up for short periods. Nonetheless, this scheme empirically outperforms native queue-based mutual exclusion implementations in most cases, sometimes drastically so. <H2>Options for enhanced scalability</h2> Version 6.0 of the collector adds two facilities to enhance collector scalability on multiprocessors. As of 6.0alpha1, these are supported only under Linux on X86 and IA64 processors, though ports to other otherwise supported Pthreads platforms should be straightforward. They are intended to be used together. <UL> <LI> Building the collector with <TT>-DPARALLEL_MARK</tt> allows the collector to run the mark phase in parallel in multiple threads, and thus on multiple processors. The mark phase typically consumes the large majority of the collection time. Thus this largely parallelizes the garbage collector itself, though not the allocation process. Currently the marking is performed by the thread that triggered the collection, together with <I>N</i>-1 dedicated threads, where <I>N</i> is the number of processors detected by the collector. The dedicated threads are created once at initialization time. <P> A second effect of this flag is to switch to a more concurrent implementation of <TT>GC_malloc_many</tt>, so that free lists can be built, and memory can be cleared, by more than one thread concurrently. <LI> Building the collector with -DTHREAD_LOCAL_ALLOC adds support for thread local allocation. It does not, by itself, cause thread local allocation to be used. It simply allows the use of the interface in <TT>gc_local_alloc.h</tt>. <P> Memory returned from thread-local allocators is completely interchangeable with that returned by the standard allocators. It may be used by other threads. The only difference is that, if the thread allocates enough memory of a certain kind, it will build a thread-local free list for objects of that kind, and allocate from that. This greatly reduces locking. The thread-local free lists are refilled using <TT>GC_malloc_many</tt>. <P> An important side effect of this flag is to replace the default spin-then-sleep lock to be replace by a spin-then-queue based implementation. This <I>reduces performance</i> for the standard allocation functions, though it usually improves performance when thread-local allocation is used heavily, and thus the number of short-duration lock acquisitions is greatly reduced. </ul> <P> The easiest way to switch an application to thread-local allocation is to <OL> <LI> Define the macro <TT>GC_REDIRECT_TO_LOCAL</tt>, and then include the <TT>gc.h</tt> header in each client source file. <LI> Invoke <TT>GC_thr_init()</tt> before any allocation. <LI> Allocate using <TT>GC_MALLOC</tt>, <TT>GC_MALLOC_ATOMIC</tt>, and/or <TT>GC_GCJ_MALLOC</tt>. </ol> <H2>The Parallel Marking Algorithm</h2> We use an algorithm similar to <A HREF="http://www.yl.is.s.u-tokyo.ac.jp/gc/">that developed by Endo, Taura, and Yonezawa</a> at the University of Tokyo. However, the data structures and implementation are different, and represent a smaller change to the original collector source, probably at the expense of extreme scalability. Some of the refinements they suggest, <I>e.g.</i> splitting large objects, were also incorporated into out approach. <P> The global mark stack is transformed into a global work queue. Unlike the usual case, it never shrinks during a mark phase. The mark threads remove objects from the queue by copying them to a local mark stack and changing the global descriptor to zero, indicating that there is no more work to be done for this entry. This removal is done with no synchronization. Thus it is possible for more than one worker to remove the same entry, resulting in some work duplication. <P> The global work queue grows only if a marker thread decides to return some of its local mark stack to the global one. This is done if the global queue appears to be running low, or if the local stack is in danger of overflowing. It does require synchronization, but should be relatively rare. <P> The sequential marking code is reused to process local mark stacks. Hence the amount of additional code required for parallel marking is minimal. <P> It should be possible to use generational collection in the presence of the parallel collector, by calling <TT>GC_enable_incremental()</tt>. This does not result in fully incremental collection, since parallel mark phases cannot currently be interrupted, and doing so may be too expensive. <P> Gcj-style mark descriptors do not currently mix with the combination of local allocation and incremental collection. They should work correctly with one or the other, but not both. <P> The number of marker threads is set on startup to the number of available processors (or to the value of the <TT>GC_NPROCS</tt> environment variable). If only a single processor is detected, parallel marking is disabled. <P> Note that setting GC_NPROCS to 1 also causes some lock acquisitions inside the collector to immediately yield the processor instead of busy waiting first. In the case of a multiprocessor and a client with multiple simultaneously runnable threads, this may have disastrous performance consequences (e.g. a factor of 10 slowdown). <H2>Performance</h2> We conducted some simple experiments with a version of <A HREF="gc_bench.html">our GC benchmark</a> that was slightly modified to run multiple concurrent client threads in the same address space. Each client thread does the same work as the original benchmark, but they share a heap. This benchmark involves very little work outside of memory allocation. This was run with GC 6.0alpha3 on a dual processor Pentium III/500 machine under Linux 2.2.12. <P> Running with a thread-unsafe collector, the benchmark ran in 9 seconds. With the simple thread-safe collector, built with <TT>-DLINUX_THREADS</tt>, the execution time increased to 10.3 seconds, or 23.5 elapsed seconds with two clients. (The times for the <TT>malloc</tt>/i<TT>free</tt> version with glibc <TT>malloc</tt> are 10.51 (standard library, pthreads not linked), 20.90 (one thread, pthreads linked), and 24.55 seconds respectively. The benchmark favors a garbage collector, since most objects are small.) <P> The following table gives execution times for the collector built with parallel marking and thread-local allocation support (<TT>-DGC_LINUX_THREADS -DPARALLEL_MARK -DTHREAD_LOCAL_ALLOC</tt>). We tested the client using either one or two marker threads, and running one or two client threads. Note that the client uses thread local allocation exclusively. With -DTHREAD_LOCAL_ALLOC the collector switches to a locking strategy that is better tuned to less frequent lock acquisition. The standard allocation primitives thus peform slightly worse than without -DTHREAD_LOCAL_ALLOC, and should be avoided in time-critical code. <P> (The results using <TT>pthread_mutex_lock</tt> directly for allocation locking would have been worse still, at least for older versions of linuxthreads. With THREAD_LOCAL_ALLOC, we first repeatedly try to acquire the lock with pthread_mutex_try_lock(), busy_waiting between attempts. After a fixed number of attempts, we use pthread_mutex_lock().) <P> These measurements do not use incremental collection, nor was prefetching enabled in the marker. We used the C version of the benchmark. All measurements are in elapsed seconds on an unloaded machine. <P> <TABLE BORDER ALIGN="CENTER"> <TR><TH>Number of threads</th><TH>1 marker thread (secs.)</th> <TH>2 marker threads (secs.)</th></tr> <TR><TD>1 client</td><TD ALIGN="CENTER">10.45</td><TD ALIGN="CENTER">7.85</td> <TR><TD>2 clients</td><TD ALIGN="CENTER">19.95</td><TD ALIGN="CENTER">12.3</td> </table> <PP> The execution time for the single threaded case is slightly worse than with simple locking. However, even the single-threaded benchmark runs faster than even the thread-unsafe version if a second processor is available. The execution time for two clients with thread local allocation time is only 1.4 times the sequential execution time for a single thread in a thread-unsafe environment, even though it involves twice the client work. That represents close to a factor of 2 improvement over the 2 client case with the old collector. The old collector clearly still suffered from some contention overhead, in spite of the fact that the locking scheme had been fairly well tuned. <P> Full linear speedup (i.e. the same execution time for 1 client on one processor as 2 clients on 2 processors) is probably not achievable on this kind of hardware even with such a small number of processors, since the memory system is a major constraint for the garbage collector, the processors usually share a single memory bus, and thus the aggregate memory bandwidth does not increase in proportion to the number of processors. <P> These results are likely to be very sensitive to both hardware and OS issues. Preliminary experiments with an older Pentium Pro machine running an older kernel were far less encouraging. </body> </html>
Go to most recent revision | Compare with Previous | Blame | View Log