StarWind Software

Hi,

I currently just network shares from one server.
I grabbed the trial version to see if i could gain any performance and so far it seems like it's worse.

I'm wondering if I'm doing something wrong.

I have a server running win2k3 with 2 raid controllers, each with 12 arrays, raid 5 gave me 2 x 3.6 TB each. I install the iscsi target and then i add a Disk Bridge Device and select one of the 3.6 TB volumes.

I have another win2k3 server run the iscsi initiator and add the target volume. i share that to the rest of my network.

I start having 3 computers transfer files to that one shared drive. speeds on each are (6-8MB/s). Without iScsi, i transferred files at (12-15MB/s )

Hardware wise, I have 2 x GigNet Switches

[iscsi target server ] => [GigNet] <=> [init server] <=> [GigNet] <=> [3 workstations]

not using MPIO. the Server with the initiator has a connection to each GigE Switch.

Thanks in advance.

Hi,

1) Please make sure network itself is capable of doing TCP at wire speed.
You can use for example NTTtcp and IPerf tools to measure maximum TCP bandwidth. You should get ~950 Mbps with this tools.
2) Optimize your TCP/IP stack to make sure it runs at full speed.
http://www.starwindsoftware.com/forums/ ... .php?t=792
3) After network itself does TCP at wire speed you need to use StarPort or MS iSCSI initiator as initiator side and StarWind with exported RAM disk to make sure SAN itself does wire speed with I/O meter or say big file copy (or whatever you use as a benchmark).
4) After you know SAN can do wire speed with low latency RAM disk you can create ImageFile target with proper settings (flat mode, async I/O and buffering enabled) or DiskBridge target and experiment with your hard disks.

Apparantly, i can't get TCP going at wirespeed....

i used iperf and did a large file transfer... i got ~300Mbps.

I'll have to look at how to improve that first.

Thanks

I optimized the server to the given settings on your FAQ...

i only get ~800Mbps when i specify a windows size (ie 300k) with iperf on the client

is that correct? or is it suppose to run at ~800 with out window size specifications on iperf?

Did you enable Jumbo frames? Did you use the same NICs on the both sides? Did you eliminate switch latency (used cross-over cable at least for the experiments)? What NTTcp tells you?