3960 SAN I/O problem.

Question

Hi all,
We have a 3960 with 2 arrays, 4 bricks with 73G drives.
It is connected to a Sun 890 for Oracle database with 2GB HBAs.
Currently, each brick is configured raid0+1 (mirror/stripe) with disk #9 as hotspare.
Disk1,2,3,4 = LUN1 ~136G
Disk 5,6,7,8 = LUN2 ~136G
Is each LUN a separate I/O channel?
Currently, we have each 136G as one single partition.
LUN1 /db01
LUN2 /db02
Do we get more performance by create 2 partitions per LUN?
LUN1 /db01a 68G and /db01b 68G
LUN2 /db02a 68G and /db02b 68G
With 1 LUN per 136G, we seem to hit a bottleneck on throughput wiht mkfile tests.
This are the Mkfile tests
time /usr/sbin/mkfile 512m TESTFILE
real 0m24.84s user 0m0.01s sys 0m2.99s
Is this slow for the T3?
Thank you for your help.
JohnHo

Answer 1

I concur that mkfile is not your best choice for a load generator. It gives you no control over of I/O size or number of threads, and also, I understand it allocates blocks from the end of the file forward, which must make sense for swap file usage but is kind of goofy for normal purposes.
I also concur that performance is context sensitive. You can not take a measurement of say 11 threads of 8 KB random write and use it to tell how fast 1 thread of 1 MB sequential read is going to be. That���s an extreme example, but you get the idea.
Using RAID-1+0 on top of RAID-5 is a very high availability approach; it is double protection against failure. For most purposes, we only design for no single point of failure. If your workload is read intensive it is also a very high performance approach. However, if you have an update intensive application, especially with small I/O, it is not your best choice for performance. It can be fine, if your I/O workload is not too sustained, but if you have sustained small writes most of time, and want maximum write performance, RAID-1+0 is a better choice. You can still use host based RAID-0 to combine LUNS and make smaller partitions if you like. If you have large sequential writes RAID-5 does pretty good if you get full stripe writes, but you need to make sure your I/O is large enough to fill a stripe (the number of data drives times the RAID stripe unit).
The goal is always to spread the load as evenly as possible over the available LUNS. If the database schema is already doing a good job of load distribution over the datafiles, then you are better off not partitioning the LUNS as it will only increase contention. On the other hand, if the load is not evenly distributed over the datafiles, combining the LUNS and partitioning them will serve you well.
There are tools available to take all the mystery out of this if you want to spend the time to design the configuration for a specific SLA given the workload composition and the capability of the chosen array configuration.
HTH,
Dave