Re: NFS and fsfs: slow commit performance

On Jan 18, 2011, at 1:49 AM, Bolstridge, Andrew wrote:

> Hi.
>
> Just a quick question/suggestion:
>
> If the exponential sleep can be a maximum of 25ms, and the random one
> between 0 and 100ms, what is the performance like if you just sleep for
> 25ms - without any randomness, or doubling of sleep times?
>
> I used to find (but this was ages ago) that just sleeping for 1ms would
> force the CPU to carry on with other things, perhaps that would have an
> effect that's just as good as the longer sleeps?
>
> Top stuff BTW, I doubt it'll have much effect on my slow virtual machine
> disks, but every performance boost helps!
> Thanks.

[cc'ing dev_at_s.a.o]

If you're not storing your repositories on NFS, then this patch doesn't help performance, the kernel's fcntl() will block the caller and unblock it when the lock is released.

The sleeps are there to not pound the NFS server with traffic.

I pt in a fixed 25ms sleep and here are the results, which don't look as good as the random sleep:

N is 1567
sum is 806.845106840133
mean is 0.514897962246415
SD N is 0.591845693041212
SD N-1 is 0.592034630218584
min is 0.1712911129
25% is 0.232829093933
50% is 0.273239135742
90% is 1.02276110649
95% is 1.61777615547
99% is 3.28569889069
max is 6.39945292473

Percentage of commits by slow systems: 27.1
Percentage of commits by fast systems: 72.9

Blair
Received on 2011-01-18 19:06:31 CET