How We Test Networked Storage Devices - Revision 3

[hua_qiu]

Just a silly questions in calculating the result of IOZone.

Both test -i 0 and -i 1 output two results, read & re-read, write & re-write. Can anyone tell me how the final results are calculated, pick a higher value or average? Thanks

 

[thiggins]

Re-read and re-write results are not used, only read and write.

 

[Fxx]

IOzone record sizes appropriate for Linux video playback?

Mr. Higgins,

As you wrote in 2006:

According to Don [Capps, creator of iozone], there are only a few record sizes that are of interest to most Windows applications: 4 K, 8 K and 64 K. 4 K is the memory page size, 8 K is one value that Windows appears to use for network transfers, and 64 K is the typical record size that Windows uses when applications try to transfer blocks of data that are bigger than 64 K.

which lead to the choice of 64K as the record size for your benchmarks.

I'm doing some testing of an external drive using different file systems and cluster sizes and would like to know your opinion about the right way to go about choosing the record size and file size.

The drive will be used mainly to store video files for playback (typically 300MB or larger), mainly but not exclusively on a Linux (Ubuntu 9.04) PC (using VLC and MPlayer, if that makes any difference).

Because the PC has 4GB RAM, I was thinking that the test filesize should be 8GB or more to mitigate the effects of caching; however, there does not seem to be very much difference in the results for the read rate performance between the 4G and 8G filesize tests.

What has me puzzled is the appropriate record size I should be using for the tests... a Google search of 'linux record size' did not reveal the answer. Also, after running a sweep of different record sizes for a 4G filesize, I noticed a disturbing anomaly -- the random read performance was significantly higher than the sequential read performance for record sizes of 512K and 1024K!

To summarize, I would appreciate your feedback on the appropriate record size to use for video playback under Linux. Thank you.

 

[thiggins]

I can't provide much guidance on this. But I will say that reads and writes to attached drives are very different than to networked shares.

 

[Fxx]

Thank you -- a follow up

I can't provide much guidance on this. But I will say that reads and writes to attached drives are very different than to networked shares.

Thank you. Yes, that makes sense...

I did find another iozone test result that exhibited the characteristic of the random read rate exceeding the sequential read rate for large record sizes, so perhaps it is not as anomalous a behavior as I had thought:

http://marc.info/?l=redhat-linux-cluster&m=111337395629359