Messages in this thread | | | Date | Fri, 31 Jan 2014 14:24:33 -0500 | From | Waiman Long <> | Subject | Re: [PATCH v3 1/2] qspinlock: Introducing a 4-byte queue spinlock implementation |
| |
On 01/31/2014 10:08 AM, Peter Zijlstra wrote: > On Tue, Jan 28, 2014 at 01:19:10PM -0500, Waiman Long wrote: >> For single-thread performance (no contention), a 256K lock/unlock >> loop was run on a 2.4Ghz Westmere x86-64 CPU. The following table >> shows the average time (in ns) for a single lock/unlock sequence >> (including the looping and timing overhead): >> >> Lock Type Time (ns) >> --------- --------- >> Ticket spinlock 14.1 >> Queue spinlock (Normal) 8.8* > What CONFIG_NR_CPUS ?
I was testing on a RHEL6.4 system which has a CONFIG_NR_CPUS of 4096.
> > Because for CONFIG_NR_CPUS< 128 (or 256 if you got !PARAVIRT), the fast > path code should be: > > ticket: > > mov $0x100,eax > lock xadd %ax,(%rbx) > cmp %al,%ah > jne ... > > although my GCC is being silly and writes: > > mov $0x100,eax > lock xadd %ax,(%rbx) > movzbl %ah,%edx > cmp %al,%dl > jne ... > > Which seems rather like a waste of a perfectly good cycle. > > With a bigger NR_CPUS you do indeed need more ops: > > mov $0x10000,%edx > lock xadd %edx,(%rbx) > mov %edx,%ecx > shr $0x10,%ecx > cmp %dx,%cx > jne ... > > > Whereas for the straight cmpxchg() you'd get something relatively simple > like: > > mov %edx,%eax > lock cmpxchg %ecx,(%rbx) > cmp %edx,%eax > jne ...
I believe the speeds of the lock functions are about the same. However, qspinlock has a much simpler unlock function which probably account of most of the speed gain.
> Anyway, as soon as you get some (light) contention you're going to tank > because you have to pull in extra cachelines, which is sad.
Light contention is the only case where the qspinlock may not perform as good as the ticket spinlock. I know this is the most common case. However, I would argue that the slowdown, if any, will not be really noticeable. This is what I will try to find out.
> I suppose we could from the ticket code more and optimize the > uncontended path, but that'll make the contended path more expensive > again, although probably not as bad as hitting a new cacheline.
I don't get what you are trying to say.
Right now, I am using only bit 0 as a lock bit. I can use bit 4, for instance, as a pending locker bit and spin until bit 0 is clear. So if there is only 1 other task spinning, it won't need to fetch another cacheline. However, it will slow down the uncontended path as I can't assign a 0 byte to free the lock. I have to use an atomic subtraction or clear bit instead.
-Longman
| |