[Lustre-discuss] max write speed of QDR HCA card

Erik Froese erik.froese at gmail.com
Thu Dec 31 09:12:40 PST 2009


Are you sure the IB card is connected at QDR rate?

ibstat will tell you. Look at the "Rate" line.

What kind of machines are you using?

ibstat
CA 'mlx4_0'
        CA type: MT26428
        Number of ports: 2
        Firmware version: 2.6.0
        Hardware version: a0
        Node GUID: 0x00212800013e5432
        System image GUID: 0x00212800013e5435
        Port 1:
                State: Active
                Physical state: LinkUp
                Rate: 40
                Base lid: 13
                LMC: 0
                SM lid: 2
                Capability mask: 0x02510868
                Port GUID: 0x00212800013e5433
        Port 2:
                State: Active
                Physical state: LinkUp
                Rate: 40
                Base lid: 14
                LMC: 0
                SM lid: 2
                Capability mask: 0x02510868
                Port GUID: 0x00212800013e5434


On Thu, Dec 31, 2009 at 2:04 AM, lakshmana swamy <klakshman03 at hotmail.com>wrote:

>  ThanQ Atul
>
> I have done the following test, According to the below test, it reaching
> 1869.74 MB/sec.
>
> This is the command I used for benchmarking.
>
> # mpiexec -machinefile machs1 -np 1024 IOR -N 1024 -a MPIIO -t 1024K -b 1G
> -F -v -o /mnt/file1
>
> Max Write: 421.31 MiB/sec (441.77 MB/sec)
> Max Read:  1029.68 MiB/sec (1079.70 MB/sec)
>
> ***
> 1. In the aboove command I used MPIIO with 1024 process, I got the almost
> same result when I fired same command with 512 process
>
> 2. I tested with stripe and without stripe with both the apis MPIIO and
> POSIX
>
>
> **
> node0:~ # ib_write_bw
> ------------------------------------------------------------------
>                     RDMA_Write BW Test
> Number of qp's running 1
> Connection type : RC
> Each Qp will post up to 100 messages each time
> Inline data is used up to 1 bytes message
>   local address:  LID 0x100, QPN 0x28004e, PSN 0xc17c83 RKey 0x8002d00
> VAddr 0x002b5fd9964000
>   remote address: LID 0x101, QPN 0x240051, PSN 0x5e61e, RKey 0x8002d00
> VAddr 0x002ae797667000
> Mtu : 2048
> ------------------------------------------------------------------
>  #bytes #iterations    BW peak[MB/sec]    BW average[MB/sec]
> node0:~ #
> node1:~ # ib_write_bw node0
> ------------------------------------------------------------------
>                     RDMA_Write BW Test
> Number of qp's running 1
> Connection type : RC
> Each Qp will post up to 100 messages each time
> Inline data is used up to 1 bytes message
>   local address:  LID 0x101, QPN 0x240051, PSN 0x5e61e RKey 0x8002d00 VAddr
> 0x002ae797667000
>   remote address: LID 0x100, QPN 0x28004e, PSN 0xc17c83, RKey 0x8002d00
> VAddr 0x002b5fd9964000
> Mtu : 2048
> ------------------------------------------------------------------
>  #bytes #iterations    BW peak[MB/sec]    BW average[MB/sec]
>   65536        5000            1869.76               1869.74
> ------------------------------------------------------------------
> node1:~ #
>
>
>
>
> ------------------------------
> Date: Thu, 31 Dec 2009 17:43:44 +1100
> From: Atul.Vidwansa at Sun.COM
> Subject: Re: [Lustre-discuss] max write speed of QDR HCA card
> To: klakshman03 at hotmail.com
> CC: lustre-discuss at lists.lustre.org
>
>
> Hi Lakshmana,
>
> You can use standard ofed tools like ib_rdma_bw to check Infiniband
> bandwidth between pair of nodes.
>
> Lustre provides lnet selftest tool for measuring IB bandwidth between pair
> or group of nodes. Check lustre manual for how tos.
>
> Finally, can you post IOR parameters you are using?
>
> Cheers,
> Atul
>
>
>
> Sent from my iPhone
>
> On Dec 31, 2009, at 5:13 PM, lakshmana swamy <klakshman03 at hotmail.com>
> wrote:
>
>
>   Dear All,
>
>   Whats the maximum write speed can be achieved with Infiniband QDR HCA
> card.
>
>
>  The Question arised from my lustre  setup where Iam not able to exceed
> 500MB/s (write speed) with two OSSs and one MDS
>
>  Each OSS has four OSTs and one HCA card. I would like to know which is the
> bottleneck here.
>
>
>  I have been using IOR for benchmarking
>
> Thank you
>
> LakshmaN
>
> ------------------------------
> New Windows 7: Simplify what you do everyday. Find the right PC for you.<http://windows.microsoft.com/shop>
>
>  _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-discuss
>
>
> ------------------------------
> New Windows 7: Find the right PC for you. Learn more.<http://windows.microsoft.com/shop>
>
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-discuss
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20091231/7f19330d/attachment.htm>


More information about the lustre-discuss mailing list