[ewg] Lustre + MPI traffic congestion free

Atul Yadav atulyadavtech at gmail.com
Tue Apr 15 00:23:17 PDT 2014


HI,

Thank for your response.
When we writing high throughput data on lustre with single switch
connectivity all things are working fine.
But when we use ftree topology for writing the same of data on lustre, the
connectivity between lustre nodes and compute nodes lost. Due to that
compute node halted.

So, can you guide us for verifying the performance of FTREE with 5 switches.
And i have no idea of"--cn_guid_file" and "--io_guid_file" using in ftree.

Thank You
Atul Yadav


On Tue, Apr 15, 2014 at 10:01 AM, Jens Domke <domke.j.aa at m.titech.ac.jp>wrote:

> Dear Atul,
>
> I'm not entirely sure what you mean with "Lustre traffic work without any
> congestion", but AFAIK there are 2 I/O-aware routing algorithms.
>
> Your can use the OpenSM flags "--cn_guid_file" and "--io_guid_file" for
> ftree (and for DFSSSP; starting from OpenSM V3.3.17) to specify the IO
> nodes and compute nodes.
>
> The purpose of io_guid_file for ftree is a bit different from what you are
> trying to accomplish according to the documentation, but I checked your
> configuration with ibsim and it works (to remove overlapping paths towards
> I/O nodes).
>
> However, if you want to separate MPI and I/O traffic completely (meaning
> no common link for both types), then this might be an impossible task with
> the currently implemented routing algorithms (but maybe Hal knows more).
>
> Regards,
> Jens
>
>
> On 15.04.14 01:52, Atul Yadav wrote:
>
>> Deal All
>>
>> We are trying to run  lustre + MPI traffic on common infiniband.
>> I am  sharing the full details of the cluster with the purpose.
>>
>> Lustre
>>
>>     - mds1
>>     - mds2
>>     - oss1
>>     - oss2
>>
>> Compute Node
>>
>>     - Nalanda
>>     - compute-0-1 to compute-0-34
>>
>>
>> Topology
>> Ftree is configured with the help of yours. 5 switch
>>
>> So, we are using common infiniband cable for Lustre and MPI traffic.
>>
>> Can i make sure my Lustre traffic work without any congestion.
>>
>> Guide us please...
>>
>> Thanks in advance
>> Atul yadav
>>
>>
>>
>> This body part will be downloaded on demand.
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/ewg/attachments/20140415/a756231e/attachment.html>


More information about the ewg mailing list