iSCSI Performance?

janderson13 · Fri Jan 14, 2011 8:06 pm

Hello,

I have installed the trial version on an IBM X-Series with a 3ghz Pentium D processor, 2GB of ram, 1GB Broadcom NIC, and 4x500GB Sata drives (lets call this BOX-A). The opertating systems is Win2008R2. When I run HDTune against the local RAID 5 array I get about 100MB/sec write and 135MB/sec read. Now... I have a 2nd X-Series server with similar specs also running Win2008R2 (lets call this BOX-B). BOX-B is accessing an iSCSI target on BOX-A. When I run HDTune against the iSCSI volume I get about 41MB/sec write and 43MB/sec read. Is that what I should expect out of iSCSI? That seems like a pretty substantial performance hit for iSCSI (which might be fine for some folks, but if you have to share the 43MB/sec between a couple of hosts that gets slow pretty quick). Any help is greatly appreciated.

Jeff

Sat Jan 15, 2011 7:33 pm

This topic has been discussed here millions of times so please try using forum help. In a nutshell:

1) It's a very BAD idea to start checking disk I/O performance w/o knowing what raw Ethernet in general and TCP in particular can push through the wire (...and how fast). No close to wire speed with TCP = no iSCSI performance.

2) HDTune is a very jerky tool as it checks linear read with one-by-one requests (something you'll NEVER do in the real life unless you have very very specific task). Please use Intel I/O Meter with a test pattern matching your mainstream task.

You should see wire speed with I/O Meter and in the real life (if you'd manage to load the wire, no "pulsating" traffic). Does not matter how fast disk I/O subsystem is - StarWind uses HUGE write-back cache to accelerate I/O on the server node.

janderson13 wrote:Hello,

I have installed the trial version on an IBM X-Series with a 3ghz Pentium D processor, 2GB of ram, 1GB Broadcom NIC, and 4x500GB Sata drives (lets call this BOX-A). The opertating systems is Win2008R2. When I run HDTune against the local RAID 5 array I get about 100MB/sec write and 135MB/sec read. Now... I have a 2nd X-Series server with similar specs also running Win2008R2 (lets call this BOX-B). BOX-B is accessing an iSCSI target on BOX-A. When I run HDTune against the iSCSI volume I get about 41MB/sec write and 43MB/sec read. Is that what I should expect out of iSCSI? That seems like a pretty substantial performance hit for iSCSI (which might be fine for some folks, but if you have to share the 43MB/sec between a couple of hosts that gets slow pretty quick). Any help is greatly appreciated.

Jeff

janderson13 · Sun Jan 16, 2011 3:25 am

Anton,

Thanks for the info - I appreciate the help. Are there any step-by-step guides for IOMeter to test the network and disk? I have messed with it a little, but since I'm not a storage guru I hate to spend a ton of time learning it for this one instance. Also, do you know if anyone is keeping a performance chart of various hardware combinations so I can see if my hardware will perform at a decent level? If not, no biggie I'll start doing my own tests. Thanks again.

Sun Jan 16, 2011 1:51 pm

OK, I see... Let's do it a bit other way: you'll tune your network to drive TCP packets with wire speed (yourself or with the help of our QA - does not matter, as you wish), you'll describe your load scenario (either here or in private - does not matter, as you wish) and we'll pick up I/O Meter test pattern for you to simulate what you're doing in the real life. Sounds good?

janderson13 wrote:Anton,

Thanks for the info - I appreciate the help. Are there any step-by-step guides for IOMeter to test the network and disk? I have messed with it a little, but since I'm not a storage guru I hate to spend a ton of time learning it for this one instance. Also, do you know if anyone is keeping a performance chart of various hardware combinations so I can see if my hardware will perform at a decent level? If not, no biggie I'll start doing my own tests. Thanks again.