Smart Array 5i write performance problem

[Juan Carlos Ramirez]

I have a write performance problem with a Smart Array 5i in a Proliant DL 380. We have write intensive real time software and we need an excellent I/O performance.

We have a Proliant DL 380 and we got a max of 83 records/sec while in a COMPAQ Evo with a 40 GB IDE HDD we got more than 500 records/sec.

Both systems have Windows2000 SP4 and SQLServer 2000 SP3. The Proliant Server has a Smart Array 5i RAID controller.

I've run the NBENCH.EXE disk benchmark in the server and my desktop COMPAQ Evo PC. We got the following results:

Compaq Evo:
file size: 999.0 MB
write: 40.24
read: 41.84

PROLIANT DL 380 Server:
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 would appreciate if somebody help me to resolve this problem

[Allen L. Clark]

I too have noted poor disk I/O performance which I believe is a hardware architecture design problem. Typically the Proliant servers have a multifunction I/O card that contains dual UltraWide SCSI channels, 10/100 NIC, and video controller all on a single card in a single motherboard slot. It is my belief that this creates a severe bandwidth bottleneck between these three data intensive devices and system RAM located on the motherboard as the data traffic for the three devices must pass thru the single I/O connector. Traffic moves slow in rush hour and with this design it is always rush hour.

I wonder if one disabled the NIC on the multifunction card and installed a separte NIC to assume the role of the disabled NIC would performance increase.

[Terrence Holt]

We ran into the same situation with a DL380G3 with a 14 disk cpq array attached. RAID-5 Read performance is great ~25MB/s, write performance to these drives is bad ~5.5MB/s. I performed several copy tests on the local fileystems and then tried the same tests on a 2 year old Dell server with a Raid controller with 64MB of write cache and slower disks. Write performance went was at least 3x better on the Dell.

We ordered a Smart Array 642 with the extra 64MB BBWC so that we can cache the disk writes and get over our hurdle.

Ben

[Doug de Werd]

The Smart 5i is really intended to provide basic RAID functionality for local server drives. The base 5i does not have a write cache, so obviously write performance will be lower. The 5i+ does have a write cache, so you will get better performance.

The tpye of RAID configuration that you have will also affect read/write performance. RAID 5 in particular is not the best to use for write intensive apps, and is made even worse with RAID 5 configurations on a controller without a write cache (the write cache is used to speed up the RAID 5 parity calculations).

When comparing benchmarks to your EVO desktop, it is not surprising that it performs well, but at the drawback that it doesn't have any availability features (like RAID).

You might want to check out this technical paper regarding RAID configurations and performance. It's a little old, but has great information.
ftp://ftp.compaq.com/pub/supportinformation/papers/ecg0110598.pdf


Here is a newer one:
http://activeanswers.compaq.com/aa_downloads/6/100/225/1/48150.pdf

Hope this helps,
Doug

[Juan Carlos Ramirez]

We know RAID 5 is not the best configuration for write intensive apps. We have four 36 GB ULTRA3 SCSI HDDs in two RAID 1 arrays.

For the performance test and the poor write performance we delete the second RAID 1 array and we create a new RAID 0 with only one of these two disks.

We don't understand why the Evo IDE 40 GB HDD has better performance than our ULTRA3 SCSI disks.

Both systems have Windows2000 SP4 and SQLServer 2000 SP3. The Proliant DL380 G3 has a Smart Array 5i RAID controller.

We update the firmware and driver version for the Smart Array controller and the results are the same.

[Terrence Holt]

As discussed:

Raid-5 parity causes overhead which means writing to a RAID-5 is usually slower than writing to a Single disk (your Evo).

Get the battery backed write cache and the writes will become much quicker because the controller will tell the OS the write is done even though it's not physically on the disk yet. When the controller is ready it will post the writes to the disks at a very fast speed.

Per Compaq, you can't add a 5i+ to a system with an integrated 5i because it's not possible to turn off the embedded 5i.

Ben

[Gert Luyten]

Per Terrence, yes you can replace the embedded SA-5i with a new SA-5iPlus & write cache since the SA-5i in the DL380-G2 is an add-on board on top of the motherboard that can be removed.
I've done this many times simply not having to explain to our customers the 'bad' write performance with RAID5 compared to a competitive server with a high-end controller.

Trust me, follow Doug's advice, he's the Compaq/HP expert!!

[Adam Thomas_2 · Accepted solution]

We have struck the same thing... I believe that it is not only a SmartArray controller shortcoming, it is also due to the nature of SCSi disks on write intensive operations.

We have a development "server" - a Compaq DeskPro PIII, 512MB RAM, standard IDE disk.

We also have a Compaq ProLiant 1850R (basically a DL380 G1) with 1GB RAM, a SmartArray 221 controller and four Ultra2 SCSI disks.

One of our developers has written a series of SQL Server stored procedures to transform data from our financial system database into fact tables in a separate database for a data warehouse.

On the DeskPro the transformation takes a fraction under an hour.

The ProLiant server takes closer to 12 hours.

A Compaq Evo D510C desktop takes 30 minutes!

I ran NBENCH and recorded very similar differences between the desktop and servers. Or DL380G3 gets nowhere near an Evo D510C for write performance, but does have the upper hand on write (36GB 15K ULTRA3 SCSI RAID 0+1)

RAID5 on a DL380G1 has absolutely appalling write performance!

Aside from the SmartArray card choking on the large write throughput with parity calculations, it seems SCSI disks are inherently bad at large sequential write operations.

In single disk configurations IDE disks are 40% faster at sequential write operations. Interleaved writes are marginally faster. SCSI is 33% faster at random writes.

There is a very good paper "Performance comparison of IDE and SCSI Disks" I found on the subject:

http://www.csl.cornell.edu/~bwhite/papers/ide_scsi.pdf

No matter how I configure my SCSI disks (on the Array Controller or on the integrated SCSI controller) I can't get the performance I need for this operation. It looks like we either buy an all singing all dancing Array Controller and 15K SCSI disks, or stick another desktop in the server room.