Обсуждение: Use gcc built-in atomic inc/dec in lock.c
Hi,
I noticed a "XXX: It might be worth considering using an atomic fetch-and-add
instruction here, on architectures where that is supported." in lock.c
Here is my first try at using it. The patch adds a configure check for
gcc 4.7 __atomic_add_fetch as well as the older __sync_add_and_fetch built-ins.
If either is available they are used instead of the mutex.
Changes do the following:
- declare atomic_inc and atomic_dec inline functions (implemented either using
GCC built-in functions or the old mutex code) in new atomics.h
- change lock.c to use atomic_* functions instead of explicit mutex
- moved all other assignments inside the mutex to occur before the atomic
operation so that the barrier of the atomic operation can guarantee the
stores are visible when the function ends
- removed one assert that could not easily be implemented with atomic_dec
in RemoveLocalLock
Using method AbortStrongLockAcquire as an example.
When compiling with Fedora GCC 4.7.2 the following assembly code is generated
Original code before the patch: 136 bytes
Mutex code with the patch: 124 bytes
Code with gcc-built-ins: 56 bytes
I think moving the extra assignments outside the mutex has allowed the
compiler to optimize the code more even when the mutex is used.
Questions:
1) is it safe to move the assignments to locallock->holdsStrongLockCount
and StrongLockInProgress outside the FastPathStrongRelationLocks?
2) With built-ins the FastPathStrongRelationLockData becomes uint32[1024],
should we add back some padding to reduce the cacheline collisions?
For a modern cpu using 64 byte cache lines there can be only
max 16 concurrent updates at the and the propability of collision
is quite big
3) What kind of pgbench test would utilise the code path the most?
TODO:
1) add check in configure.in to ensure that the built-ins are not converted
to external calls by gcc (on architectures that use the gcc generic version)
2) other architectures / compilers
------
Original code before the patch:
0000000000000e10 <AbortStrongLockAcquire>:
e10: 48 89 5c 24 f0 mov %rbx,-0x10(%rsp)
e15: 48 89 6c 24 f8 mov %rbp,-0x8(%rsp)
e1a: 48 83 ec 18 sub $0x18,%rsp
e1e: 48 8b 1d 00 00 00 00 mov 0x0(%rip),%rbx # e25 <AbortStrongLockAcquire+0x15>
e25: 48 85 db test %rbx,%rbx
e28: 74 41 je e6b <AbortStrongLockAcquire+0x5b>
e2a: 8b 6b 28 mov 0x28(%rbx),%ebp
e2d: b8 01 00 00 00 mov $0x1,%eax
e32: 48 8b 15 00 00 00 00 mov 0x0(%rip),%rdx # e39 <AbortStrongLockAcquire+0x29>
e39: 81 e5 ff 03 00 00 and $0x3ff,%ebp
e3f: f0 86 02 lock xchg %al,(%rdx)
e42: 84 c0 test %al,%al
e44: 75 3a jne e80 <AbortStrongLockAcquire+0x70>
e46: 48 8b 05 00 00 00 00 mov 0x0(%rip),%rax # e4d <AbortStrongLockAcquire+0x3d>
e4d: 48 c7 05 00 00 00 00 movq $0x0,0x0(%rip) # e58 <AbortStrongLockAcquire+0x48>
e54: 00 00 00 00
e58: 83 6c a8 04 01 subl $0x1,0x4(%rax,%rbp,4)
e5d: c6 43 40 00 movb $0x0,0x40(%rbx)
e61: 48 8b 05 00 00 00 00 mov 0x0(%rip),%rax # e68 <AbortStrongLockAcquire+0x58>
e68: c6 00 00 movb $0x0,(%rax)
e6b: 48 8b 5c 24 08 mov 0x8(%rsp),%rbx
e70: 48 8b 6c 24 10 mov 0x10(%rsp),%rbp
e75: 48 83 c4 18 add $0x18,%rsp
e79: c3 retq
e7a: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
e80: 48 8b 3d 00 00 00 00 mov 0x0(%rip),%rdi # e87 <AbortStrongLockAcquire+0x77>
e87: ba a8 05 00 00 mov $0x5a8,%edx
e8c: be 00 00 00 00 mov $0x0,%esi
e91: e8 00 00 00 00 callq e96 <AbortStrongLockAcquire+0x86>
e96: eb ae jmp e46 <AbortStrongLockAcquire+0x36>
e98: 0f 1f 84 00 00 00 00 nopl 0x0(%rax,%rax,1)
e9f: 00
Mutex code with the patch:
0000000000000e00 <AbortStrongLockAcquire>:
e00: 48 8b 05 00 00 00 00 mov 0x0(%rip),%rax # e07 <AbortStrongLockAcquire+0x7>
e07: 48 85 c0 test %rax,%rax
e0a: 74 55 je e61 <AbortStrongLockAcquire+0x61>
e0c: 48 89 5c 24 f0 mov %rbx,-0x10(%rsp)
e11: 48 89 6c 24 f8 mov %rbp,-0x8(%rsp)
e16: 48 83 ec 18 sub $0x18,%rsp
e1a: 8b 68 28 mov 0x28(%rax),%ebp
e1d: c6 40 40 00 movb $0x0,0x40(%rax)
e21: b8 01 00 00 00 mov $0x1,%eax
e26: 48 c7 05 00 00 00 00 movq $0x0,0x0(%rip) # e31 <AbortStrongLockAcquire+0x31>
e2d: 00 00 00 00
e31: 48 8b 1d 00 00 00 00 mov 0x0(%rip),%rbx # e38 <AbortStrongLockAcquire+0x38>
e38: 81 e5 ff 03 00 00 and $0x3ff,%ebp
e3e: f0 86 03 lock xchg %al,(%rbx)
e41: 84 c0 test %al,%al
e43: 75 23 jne e68 <AbortStrongLockAcquire+0x68>
e45: 8b 44 ab 04 mov 0x4(%rbx,%rbp,4),%eax
e49: 83 e8 01 sub $0x1,%eax
e4c: 89 44 ab 04 mov %eax,0x4(%rbx,%rbp,4)
e50: c6 03 00 movb $0x0,(%rbx)
e53: 48 8b 5c 24 08 mov 0x8(%rsp),%rbx
e58: 48 8b 6c 24 10 mov 0x10(%rsp),%rbp
e5d: 48 83 c4 18 add $0x18,%rsp
e61: f3 c3 repz retq
e63: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
e68: ba 41 00 00 00 mov $0x41,%edx
e6d: be 00 00 00 00 mov $0x0,%esi
e72: 48 89 df mov %rbx,%rdi
e75: e8 00 00 00 00 callq e7a <AbortStrongLockAcquire+0x7a>
e7a: eb c9 jmp e45 <AbortStrongLockAcquire+0x45>
e7c: 0f 1f 40 00 nopl 0x0(%rax)
Code with gcc-built-ins:
0000000000000da0 <AbortStrongLockAcquire>:
da0: 48 8b 05 00 00 00 00 mov 0x0(%rip),%rax # da7 <AbortStrongLockAcquire+0x7>
da7: 48 85 c0 test %rax,%rax
daa: 74 2a je dd6 <AbortStrongLockAcquire+0x36>
dac: 8b 50 28 mov 0x28(%rax),%edx
daf: c6 40 40 00 movb $0x0,0x40(%rax)
db3: 48 c7 05 00 00 00 00 movq $0x0,0x0(%rip) # dbe <AbortStrongLockAcquire+0x1e>
dba: 00 00 00 00
dbe: 48 89 d0 mov %rdx,%rax
dc1: 48 8b 15 00 00 00 00 mov 0x0(%rip),%rdx # dc8 <AbortStrongLockAcquire+0x28>
dc8: 25 ff 03 00 00 and $0x3ff,%eax
dcd: 48 c1 e0 02 shl $0x2,%rax
dd1: f0 83 2c 02 01 lock subl $0x1,(%rdx,%rax,1)
dd6: f3 c3 repz retq
dd8: 0f 1f 84 00 00 00 00 nopl 0x0(%rax,%rax,1)
ddf: 00
Вложения
On 12 December 2012 22:11, Mikko Tiihonen <mikko.tiihonen@nitorcreations.com> wrote: > noticed a "XXX: It might be worth considering using an atomic fetch-and-add > instruction here, on architectures where that is supported." in lock.c > > Here is my first try at using it. That's interesting, but I have to wonder if there is any evidence that this *is* actually helpful to performance. -- Peter Geoghegan http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training and Services
On 12/13/2012 12:19 AM, Peter Geoghegan wrote: > On 12 December 2012 22:11, Mikko Tiihonen > <mikko.tiihonen@nitorcreations.com> wrote: >> noticed a "XXX: It might be worth considering using an atomic fetch-and-add >> instruction here, on architectures where that is supported." in lock.c >> >> Here is my first try at using it. > > That's interesting, but I have to wonder if there is any evidence that > this *is* actually helpful to performance. One of my open questions listed in the original email was request for help on creating a test case that exercise the code path enough so that it any improvements can be measured. But apart from performance I think there are two other aspects to consider: 1) Code clarity: I think the lock.c code is easier to understand after the patch 2) Future possibilities: having the atomic_inc/dec generally available allows other performance critical parts of postgrestake advantage of them in the future -Mikko
On Fri, Dec 14, 2012 at 9:33 AM, Mikko Tiihonen <mikko.tiihonen@nitorcreations.com> wrote: > On 12/13/2012 12:19 AM, Peter Geoghegan wrote: >> >> On 12 December 2012 22:11, Mikko Tiihonen >> <mikko.tiihonen@nitorcreations.com> wrote: >>> >>> noticed a "XXX: It might be worth considering using an atomic >>> fetch-and-add >>> instruction here, on architectures where that is supported." in lock.c >>> >>> Here is my first try at using it. >> >> >> That's interesting, but I have to wonder if there is any evidence that >> this *is* actually helpful to performance. > > > One of my open questions listed in the original email was request for help > on > creating a test case that exercise the code path enough so that it any > improvements can be measured. > > But apart from performance I think there are two other aspects to consider: > 1) Code clarity: I think the lock.c code is easier to understand after the > patch > 2) Future possibilities: having the atomic_inc/dec generally available > allows > other performance critical parts of postgres take advantage of them in > the > future This was actually attempted a little while back; a spinlock was replaced with a few atomic increment and decrement calls for managing the refcount and other things on the freelist. It helped or hurt depending on contention but the net effect was negative. On reflection I think that was because that the assembly 'lock' instructions are really expensive relative to the others: so it's not safe to assume that say 2-3 gcc primitive increment calls are cheaper that a spinlock. merlin
On 12/14/2012 05:55 PM, Merlin Moncure wrote: > On Fri, Dec 14, 2012 at 9:33 AM, Mikko Tiihonen > <mikko.tiihonen@nitorcreations.com> wrote: >> On 12/13/2012 12:19 AM, Peter Geoghegan wrote: >>> >>> On 12 December 2012 22:11, Mikko Tiihonen >>> <mikko.tiihonen@nitorcreations.com> wrote: >>>> >>>> noticed a "XXX: It might be worth considering using an atomic >>>> fetch-and-add >>>> instruction here, on architectures where that is supported." in lock.c >>>> >>>> Here is my first try at using it. >>> >>> >>> That's interesting, but I have to wonder if there is any evidence that >>> this *is* actually helpful to performance. >> >> >> One of my open questions listed in the original email was request for help >> on >> creating a test case that exercise the code path enough so that it any >> improvements can be measured. >> >> But apart from performance I think there are two other aspects to consider: >> 1) Code clarity: I think the lock.c code is easier to understand after the >> patch >> 2) Future possibilities: having the atomic_inc/dec generally available >> allows >> other performance critical parts of postgres take advantage of them in >> the >> future > > This was actually attempted a little while back; a spinlock was > replaced with a few atomic increment and decrement calls for managing > the refcount and other things on the freelist. It helped or hurt > depending on contention but the net effect was negative. On > reflection I think that was because that the assembly 'lock' > instructions are really expensive relative to the others: so it's not > safe to assume that say 2-3 gcc primitive increment calls are cheaper > that a spinlock. The spinlock uses one 'lock' instruction when taken, and no lock instructions when released. Thus I think replacing one spinlock protected add/sub with atomic 'lock' add/sub not perform worse. But if you replace "mutex-lock,add,add,unlock" with "atomic add, atomic add" you already have more hw level synchronization and thus risk lower performance if they are on separate cache lines. This of course limits the use cases of the atomic operations. -Mikko
On Wed, Dec 12, 2012 at 5:19 PM, Peter Geoghegan <peter@2ndquadrant.com> wrote: > On 12 December 2012 22:11, Mikko Tiihonen > <mikko.tiihonen@nitorcreations.com> wrote: >> noticed a "XXX: It might be worth considering using an atomic fetch-and-add >> instruction here, on architectures where that is supported." in lock.c >> >> Here is my first try at using it. > > That's interesting, but I have to wonder if there is any evidence that > this *is* actually helpful to performance. Ditto. I've had at least one bad experience with an attempted change to this sort of thing backfiring. And it's possible that's because my implementation sucked, but the only concrete evidence of that which I was able to discern was bad performance. So I've learned to be skeptical of these kinds of things unless there are clear benchmark results. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
----- Цитат от Mikko Tiihonen (mikko.tiihonen@nitorcreations.com), на 14.12.2012 в 17:33 -----<br /><br />> On 12/13/201212:19 AM, Peter Geoghegan wrote:<br />>> On 12 December 2012 22:11, Mikko Tiihonen<br />>> wrote:<br/>>>> noticed a "XXX: It might be worth considering using an atomic fetch-and-add<br />>>> instructionhere, on architectures where that is supported." in lock.c<br />>>><br />>>> Here is my firsttry at using it.<br />>><br />>> That's interesting, but I have to wonder if there is any evidence that<br/>>> this *is* actually helpful to performance.<br />> <br />> One of my open questions listed in theoriginal email was request for help on<br />> creating a test case that exercise the code path enough so that it any<br/>> improvements can be measured.<br />> <br /><br />Running pgbench on 16+ cores/threads could stress lockingprimitives. From my experience even benchmarks run on 8 core systems should tell the difference.<br /><br />--<br/>Luben Karavelov