Performance numbers?
Eric Anderson
anderson at freebsd.org
Tue Feb 3 05:08:05 PST 2009
On Feb 2, 2009, at 3:16 AM, Ivan Voras wrote:
> Eric Anderson wrote:
>>
>> On Jan 30, 2009, at 2:50 PM, Ivan Voras wrote:
>>
>>> Eric Anderson wrote:
>>>> Hi GEOMers!
>>>>
>>>> Does anyone have any benchmarks or numbers relating to GEOM
>>>> performance?
>>>>
>>>> I tried doing some on my own, but I didn't get very satisfactory
>>>> results, so I'm curious what others have seen or used.
>>>>
>>>> My hardware is a Core 2 Quad, with 4GB of ram.
>>>>
>>>> First, I made an mdconfig'ed malloc backed 'disk' of 1.5GB.
>>>> Then, I
>>>> tried running such tools as rawio, and diskinfo. rawio fails with
>>>> input/output errors, and diskinfo wants a larger device to give
>>>> the full
>>>> stats. I ended up using purely dd since that worked.
>>>> Interestingly
>>>> enough, dd'ing to the malloc device results in about 1000
>>>> operations per
>>>> second, regardless of a blocksize of 512bytes or 1MB.
>>>
>>> It's a good idea for testing.
>>>
>>> 1000 ops/s looks suspiciously like HZ, though I don't know why HZ
>>> would
>>> influence GEOM (AFAIK context switches between threads, including
>>> GEOM
>>> threads do not depend on it) - can you try ruling out HZ?
>>
>> Is there a way to pump the data through the GEOM layers without
>> doing a
>> mdconfig'd disk?
>
> You could use gzero, it ignores written data and produces read data
> with
> memset.
Ok, I tried gzero, and now the numbers are *much* different. I'm
getting roughly 60,000 ops/s now single threaded, or about 140,000 ops/
s using 4 threads. Much better! :) I could not for the life of me
remember geom_zero, so thanks for the reminder.
>
>
>> Also, are you thinking setting the hz to some other setting, and
>> rerunning? What setting were you thinking? I can easily try
>> anything.
>
> Yes. Try 1500.
>
I think I'll skip that now that I have gzero working giving me more
realistic numbers.
Eric
More information about the freebsd-geom
mailing list