[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: file systems

On 5/2/2011 5:54 PM, Boyd Stephen Smith Jr. wrote:

I'm slightly surprised by the results.  It's possible it was slightly weighted
toward JFS because of the "%CPU" and "Ops/%CPU" metrics, which I don't think
matter too much.

As I mentioned previously, the only relevant graph of each set is the one at the top of each page, either MB/s or IOPS. The other graphs are useless for any kind of ranking, and in fact will likely produce misleading rankings. If your ranking program is looking at tall bars, it will make an inverse ranking error WRT %CPU data, where the lower bars are better. %CPU isn't relevant anyway as any 4+ core server shipped in the past 3 years has a huge excess of CPU/mem bandwidth in relation to IO.

If you rank strictly based on the first graph of each test result page you'll see why I use the word "trounces".

I'd love to see data for 2.6.32 (Squeeze) and 2.6.38 (Wheezy/Sid).

XFS had horrible metadata performance until 2.6.35 which introduced delayed logging, so the mail server test will have much lower results with the stock Squeeze kernel. There was an additional XFS patch in 2.6.39 which speeds metadata operations up even further. From 2.6.35 to 2.6.38 you won't see much, if any, difference in XFS performance. I don't keep up with the other filesystems so I can't comment on their development.


Reply to: