Peter Samuelson wrote:
> [Blair Zajac]
>> We're seeing around 10 commits per second, getting up to 20 commits
>> per second during busy parts of the day. So I'm concerned about
>> sharing the write lock for the commit process and just incrementing a
>> transaction sequence.
>>
>> I don't know how fast getting a lock, incrementing the counter and
>> releasing it will be, nor how much contention.
>
> On my 1.8 GHz Athlon, I can do this:
>
> open
> fcntl(F_WRLCK)
> read
> atoi
> lseek
> write(n+1)
> close [note that this unlocks the file]
> return n
>
> around 70000 times per second. Thus, if I wanted to do it 20 times per
> second, I wouldn't worry much about either efficiency or contention.
Hi Peter,
Nice test. Thanks for putting it together.
I'm seeing about the same speed. Funny, to our NetApps, they are only slightly
slower (70,000) than the mirrored disk (77,000). Going to non-mirrored disk
gets around 90,000 per second.
I'll stick with the single lock.
Regards,
Blair
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Fri Jul 13 00:55:45 2007