Optimal File System config for 2.5TB RAID5

Danny Do ai_quoc at hotmail.com
Tue Sep 30 13:37:21 UTC 2008

>Why do you think slower drives using an interface that has known
>problems handling concurrent connections will be faster than faster
>drives using an interface designed for concurrency?

My current 6x300GB SCSI system using the FreeBSD "default max raw I/O
transfer size" (64KB). Assume that all reads are random. In order to read
1MB from the hard drive, it would cost:
-> 1024/64 * (seek time + time to read 64K)
-> 16 * (8ms + <1ms) [average seek time on SATA 7200RPM is 8ms, make it 0ms
for read time]
-> 128ms to read 1MB

If I change the "default max raw I/O transfer size" to 1MB it would only
cost (8ms seek time + 2.6ms read 1MB using SATA300). So, the time to read
1MB is only about 10.6ms.

As we can see here reading 1MB from the hard disk is at least 10 times
better if we increase the "default max raw I/O transfer size" to 1MB. This
is mainly because the main cost for reading random data from hard disk is
seek time.

I think the drawback from such configuration is that our server will consume
at least:
- n concurrent connections * "default max raw I/O transfer size" 
of memory just for reading the data from hard disk. RAM quite cheap these
days, I think it's ok.

>Based on my experiences with SATA vs. U160/U320 SCSI or SAS your likely
>outcome is to see a marked decrease in performance.  I'd be interested
>to hear your results.

If both SATA and SCSI system using the same configuration, the end result
should be obvious. However, If SCSI system using 64K IO transfer size whilst
SATA using 1MB IO transfer size, I don't know! I think the SATA system will
outperform the SCSI system. 

I'll let you know when I get the new SATA system from my ISP.



More information about the freebsd-questions mailing list