digitalmars.D - More on GC Spinlocks
- dsimcha <dsimcha yahoo.com> Dec 06 2008
- Fawzi Mohamed <fmohamed mac.com> Dec 06 2008
- The Anh Tran <trtheanh gmail.com> Dec 06 2008
- bearophile <bearophileHUGS lycos.com> Dec 07 2008
- dsimcha <dsimcha yahoo.com> Dec 07 2008
- The Anh Tran <trtheanh gmail.com> Dec 09 2008
A few days ago, I commented that I thought that maybe the GC should be using
spinlocks, given how little time a typical allocation takes compared to
context switches, etc. I've created a version of the D 2.21 druntime GC with
spinlocks instead of synchronized, and created the following simple benchmark
to just generate a ton of contention for the GC:
import core.thread, core.memory, std.perf, std.stdio, std.c.time, std.c.stdio;
void main() {
readln; //Allow for affinity to be changed.
GC.disable;
auto T = new Thread(&foo);
T.start;
scope auto pc = new PerformanceCounter;
pc.start;
foo();
T.join;
pc.stop;
writeln(pc.milliseconds);
}
void foo() {
foreach(i; 0..10_000_000) {
auto foo = GC.malloc(8);
GC.free(foo);
}
}
Here are the times:
Using both of my CPU cores, meaning serious contention, in milliseconds:
Spinlock: 10006
Synchronized: 28563
The synchronized version uses ~25-30% CPU, because of OS rescheduling, while
the spinlock version uses 100%.
Setting the affinity to only one CPU to simulate a single-CPU environment:
Spinlock: 4356
Synchronized: 4758
Replacing one thread's foo() by a dummy function so that the lock is never
even contested:
Spinlock: 1876
Synchronized: 2589
I will acknowledge that this is an extremely simple benchmark, but I think
it's reasonably representative of a severely contested memory allocation lock.
The spinlock I used was the simplest possible atomic CAS lock, nothing fancy.
Dec 06 2008
On 2008-12-06 21:31:34 +0100, dsimcha <dsimcha yahoo.com> said:A few days ago, I commented that I thought that maybe the GC should be using spinlocks, given how little time a typical allocation takes compared to context switches, etc. I've created a version of the D 2.21 druntime GC with spinlocks instead of synchronized, and created the following simple benchmark to just generate a ton of contention for the GC: import core.thread, core.memory, std.perf, std.stdio, std.c.time, std.c.stdio; void main() { readln; //Allow for affinity to be changed. GC.disable; auto T = new Thread(&foo); T.start; scope auto pc = new PerformanceCounter; pc.start; foo(); T.join; pc.stop; writeln(pc.milliseconds); } void foo() { foreach(i; 0..10_000_000) { auto foo = GC.malloc(8); GC.free(foo); } } Here are the times: Using both of my CPU cores, meaning serious contention, in milliseconds: Spinlock: 10006 Synchronized: 28563 The synchronized version uses ~25-30% CPU, because of OS rescheduling, while the spinlock version uses 100%. Setting the affinity to only one CPU to simulate a single-CPU environment: Spinlock: 4356 Synchronized: 4758 Replacing one thread's foo() by a dummy function so that the lock is never even contested: Spinlock: 1876 Synchronized: 2589 I will acknowledge that this is an extremely simple benchmark, but I think it's reasonably representative of a severely contested memory allocation lock. The spinlock I used was the simplest possible atomic CAS lock, nothing fancy.
Nice, I think indeed for the allocation I think that spinlock are definitely preferable, but I think that they should be integrated in the runtime, so that the runtime can be built also on platforms that do not support them (maybe a spinlock module that can be implemented on the top of normal locks if the need arises, it should not add too much overhead I hope. Fawzi
Dec 06 2008
Hi, Could you measure spinlock with this bench? http://shootout.alioth.debian.org/u32q/benchmark.php?test=binarytrees&lang=all Thanks.
Dec 06 2008
bearophile <bearophileHUGS lycos.com> The Anh Tran:Could you measure spinlock with this bench? http://shootout.alioth.debian.org/u32q/benchmark.php?test=binarytrees&lang=all
This is a D version designed to be simple (it's like the Java version): http://shootout.alioth.debian.org/gp4/benchmark.php?test=binarytrees&lang=dlang&id=1 Bye, bearophile
Dec 07 2008
== Quote from The Anh Tran (trtheanh gmail.com)'s articleHi, Could you measure spinlock with this bench? http://shootout.alioth.debian.org/u32q/benchmark.php?test=binarytrees&lang=all Thanks.
I tried to reply to your post last night with my modified gcx.d, but apparently posting attachments of that size (~80kb) silently fails. There is no multithreaded implementation of the binary trees benchmark for D that I was able to find. As far as the single-threaded version, spinlocks would make absolutely no difference because the druntime GC uses thread_needLock() to avoid any kind of lock on single-threaded code. If you or anyone else wants to play around w/ my modified gcx.d code and try it under different use cases, I've posted it to http://cis.jhu.edu/~dsimcha/gcx.d.
Dec 07 2008
Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit dsimcha wrote:== Quote from The Anh Tran (trtheanh gmail.com)'s articleHi, Could you measure spinlock with this bench? http://shootout.alioth.debian.org/u32q/benchmark.php?test=binarytrees&lang=all Thanks.
I tried to reply to your post last night with my modified gcx.d, but apparently posting attachments of that size (~80kb) silently fails. There is no multithreaded implementation of the binary trees benchmark for D that I was able to find. As far as the single-threaded version, spinlocks would make absolutely no difference because the druntime GC uses thread_needLock() to avoid any kind of lock on single-threaded code. If you or anyone else wants to play around w/ my modified gcx.d code and try it under different use cases, I've posted it to http://cis.jhu.edu/~dsimcha/gcx.d.
I've failed to recompile druntime with your gcx.d. I'm still a D newbie :|. Ie: need someone to hold my hand and guide steps by steps ;) This is my multithread alloc implementation. If it uses std.gc, and runs with 2 threads; it is 5 times slower than 1 thread !???
Dec 09 2008