File Transfer Speeds

[sean.thompson]

Hi,

We are the owners of a shiny new UCS system which we plan to use to host a sql cluster, consisting of 2 chassis with 3 blades in each chassis. There is very little advanced customisation at the moment, the vNICs have the windows adapter policy applied. I have installed Windows 2008 R2 on 2 of the blades in one chassis and am confused by the data tranfer rates I am getting between blades.

If I run parallel streams using iperf, I see data transfer rates of arround 9.8gb/s. If I use a single stream, it never exceeds around 1gb/s. File transfers between blades and to netapp storage and tests using SQLIO show similar speeds or slower

I apprecate that this is a bit vague, I am really looking for some pointers as to why this would happen and what I can do to rectify it    - the bandwidth is obviously there, it is just not being utilised          

[Robert Burns]

Sean,

How are you running the tests exactly path-wise?  If you run a test between blades using a vNIC that goes to the SAME fabric interconnect, the traffic remains within UCS.  If your source blade uses a vNIC going to Fabric-A and the destination blade's target vNIC is using Fabric-B, then you're going up to Fabric Interconnect A, to a northbound switch, and back down to Fabric Interconnect B and to the destination blade.

Try to perform your tests using vNICs from Fabric-A only, then compare the results to a Fabric-A -> Fabric-B test.  This could point to a bottleneck northbound of UCS.

Also, iperf is not perfectly suited for Windows performance.  You're better off using ntttcp which is a Windows specific performance tool.

Regards,

Robert

[sean.thompson]

Thanks Robert,

The end goal is to have traffic through a software iSCSI connection to a netapp FAS, however as I am seeing the slow transfer speed just copying an iso between blades in the same chassis I thought that I would try to get that sorted first to minimise the potential bottlenecks.

I'll give NTttcp a try to see what that reports.

Regards

Sean