[Lustre-discuss] performance issues in simple lustre setup

[Murray Smigel]

Isaac Huang wrote:

>On Mon, Jun 02, 2008 at 01:40:20PM -0400, Murray Smigel wrote:
>  
>
>>Hi,
>>I have built a simple lustre setup. MDS and OSS are both on a Centos5
>>machine using the red hat lustre modified kernel
>>2.6.18-8.1.14.el5_lustre.1.6.4.1 running OFED-1.3.  Lustre is 1.6.4.3.
>>
>>The client machine is Debian running the same kernel and OFED-1.3 and
>>lustre 1.6.4.3.
>>
>>The MDT and OST are both single partitions on the the same disk (yes,
>>I know this is not optimal...)
>>The network uses Mellanox ConnectX HCAs running through a Voltaire
>>ISR2004 switch.
>>    
>>
>
>What is you Lustre network configurations (i.e. lnet options)? If not
>sure, what's the output of 'lctl list_nids' on the client and the
>server?
>
>  
>
**
On server:  (I have tried both msi_x=1 and 0 with no difference)
modprobe.conf -->
options lnet networks="tcp,o2ib"

options mlx4_core       msi_x=0
alias ib0 ib_ipoib
alias ib1 ib_ipoib

[root at lusty lustre-iokit-1.2]# lctl list_nids
192.168.1.94 at tcp
192.168.3.50 at o2ib


On client:
/etc/modprobe.d/lustre -->
options lnet networks="tcp,o2ib"

options mlx4_core       msi_x=1
alias ib0 ib_ipoib
alias ib1 ib_ipoib

nasnu3:/home/murray/lustre-iokit-1.2# lctl list_nids
192.168.1.156 at tcp
192.168.3.30 at o2ib

Thanks,
murray

>Isaac
>
>  
>
>>The basic RDMA setup seems to work in either direction:
>>murray at nasnu3:/slut$ ib_rdma_bw 192.168.3.50 (Lustre server)
>>5605: | port=18515 | ib_port=1 | size=65536 | tx_depth=100 |
>>iters=1000 | duplex=0 | cma=0 |
>>5605: Local address:  LID 0x07, QPN 0x22004e, PSN 0x323aa6 RKey
>>0x1a002800 VAddr 0x002aaaaaad6000
>>5605: Remote address: LID 0x05, QPN 0x8004f, PSN 0x67c28c, RKey
>>0x8002800 VAddr 0x002aaaab705000
>>
>>
>>5605: Bandwidth peak (#0 to #985): 1332.53 MB/sec
>>5605: Bandwidth average: 1332.47 MB/sec
>>5605: Service Demand peak (#0 to #985): 1462 cycles/KB
>>5605: Service Demand Avg  : 1462 cycles/KB
>>
>>[murray at lusty bin]$ ib_rdma_bw 192.168.3.30 (Lustre client)
>>3845: | port=18515 | ib_port=1 | size=65536 | tx_depth=100 |
>>iters=1000 | duplex=0 | cma=0 |
>>3845: Local address:  LID 0x05, QPN 0xa004f, PSN 0x4f4712 RKey
>>0xa002800 VAddr 0x002aaaab705000
>>3845: Remote address: LID 0x07, QPN 0x24004e, PSN 0xa740c1, RKey
>>0x1c002800 VAddr 0x002aaaaaad6000
>>
>>
>>3845: Bandwidth peak (#0 to #956): 1533.5 MB/sec
>>3845: Bandwidth average: 1533.43 MB/sec
>>3845: Service Demand peak (#0 to #956): 1146 cycles/KB
>>3845: Service Demand Avg  : 1146 cycles/KB
>>
>>Local disk speed on the Lustre server seems fine, as does speed when
>>the Lustre machine writes
>>to the Lustre mounted drive (50-80 MB/s).
>>[murray at lusty slut]$ dd if=/dev/zero of=foo bs=1048576 count=1024
>>1024+0 records in
>>1024+0 records out
>>1073741824 bytes (1.1 GB) copied, 13.5875 seconds, 79.0 MB/s
>>
>>Performance of the client machine writing to the Lustre drive is poor
>>(12 MB/s)
>>murray at nasnu3:/slut$ mount -t lustre -l
>>192.168.3.50 at o2ib:/lusty on /slut type lustre (rw)
>>
>>murray at nasnu3:/slut$ dd if=/dev/zero of=foo bs=1048576 count=1024
>>1024+0 records in
>>1024+0 records out
>>1073741824 bytes (1.1 GB) copied, 88.9857 seconds, 12.1 MB/s
>>
>>Similar results from using Bonnie++ for the testing.
>>
>>Any ideas as to what might be going on?
>>Thanks,
>>murray
>>
>>_______________________________________________
>>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/20080603/da1c6fad/attachment.htm>