Update for all: DL380 5i Plus performance

[Terrence Holt]

All,
Like others, we were having issues with write performance on the integrated 5i plus on the DL380 G3. We added the battery ($150) to the 5i Plus and after it charged up we have about 50MB of cache to use. I set it to 25% read/ 75% write and these are the improvements I achieved:

C: Raid1 internal
E: Raid1 on cpq storage enc.
F: 12 disk Raid 5 on cpq storage enc.

>>> Original Setup:
copy from f: to e: 8.8MB/s
copy from e: to c: 8.8MB/s
copy from c: to f: 5.4MB/s (not so good)

>> new, with battery
Copy from F: to e: 22MB/s (~250% increase)
Copy from E: to C: 24MB/s (~250% increase)
Copy from C: to F: 28MB/s (518% increase)


I had to flush the write cache before I did the write test on F: test because I was getting 34MB/s, some of the previous runs were being cached. The coolest thing is that E: is RAID1 and F: is RAID5. Normally RAID1 is faster writing than RAID5 because of the lack of parity generation. The 5i Plus controller is doing a great job of parity calculation and then bursting the writes out to the F: disk from cache!

[Juan Carlos Ramirez]

We also test it with a new SMART ARRAY 5302/128 in our PROLIANT DL380 G3. Our disk configuration for tests is:
C: RAID 1 (2 disks)
D: RAID 0 (1 disk)

We got 166.67MB/s write and 150.91MB/s read on files smaller than cache size (128 MB). If we test with files greater than cache size the results are similar to the embedded SA-5i.

embedded SA-5i:
RAID 1:
file size: 999.0 MB
write: 11.73
read: 41.98

RAID 0 36GB single disk:
file size: 999.0 MB
write: 12.61
read: 66.39

I think this is an internal performance problem in these SCSI disks, because our results in IDE disks are better.

Other thing we observed is that CPU utilization in SCSI test was ~10% while in the IDE tests was ~95-100%. This is good for SCSI but why we want a low CPU utilization if we can't get a better performance.

The real problem is that we have a write intensive real time software and we need an excellent I/O performance.

[Doug de Werd]

One of the things to keep in mind about overall performance is that benchmarks don't always tell the whole story. In regards to the comment about IDE performing better, it is significant that the CPU utilization was >90%. This means that when running your real application, there is tremendous contention for CPU resources, and in the real world, this would balance out with much lower disk throughput on IDE when running your app, and it would also probably max out the CPU usage near 100% all the time. This means that you have no additional headroom for growth (like additional users/transactions/etc.). The fact that SCSI has very low CPU usage means that your application has much more CPU resources available to it, providing more headroom and scalability.

SCSI systems such as those used by Smart Controllers also have the ability to take advantage of hardware RAID, which is another factor to consider (in addition to the much larger capacities supported via RAID configurations). Obviously a single IDE drive is not going to provide this - they are really designed for single user/single task environments. SCSI has a lot of things in there that can maintain performance with larger loads because they are optimized for multi-tasking (like disk elevator seek) and things like Dynamic Sector Repair and SMART analysis.

You might be interested in some of these links as well. They provide a very good overview of both technologies, and why/where each one fits best:

SCSI Drive Reliability
ftp://ftp.compaq.com/pub/products/servers/proliantstorage/drives-enclosures/scsi-hdd-quality-sys.pdf

SCSI vs. IDE Performance
http://www6.tomshardware.com/storage/20010129/


SCSI vs. IDE Comparison
http://www.hardwarecentral.com/hardwarecentral/tutorials/43/1/

SCSI and IDE Technical Overview
http://www.acc.umu.se/~sagge/scsi_ide/#comparison

SCSI/IDE Technical Comparison
http://www.storagereview.com/guide2000/ref/hdd/if/compCut.html
Just some things to think about...

Thanks,
Doug