Tziporet, 

This did appear to be a firmware issue. After the update I have a latency difference on the order of .2 uSec which seems reasonable, much better than a ~5. uSec penalty!  

Thanks for the help, 

Galen 
 

On Jul 27, 2005, at 12:38 AM, Tziporet Koren wrote:

Hi Galen,
You are working with an old FW version that had issues with SRQ performance.
We did performance improvements and now the latency and BW with SRQ are the same as with regular QP.
 
The changes are available in the new FW release we did few weeks ago: see http://www.mellanox.com/products/firmware.html
 
Regarding your specific device: you are using MT25208 InfiniHost III Ex, so you need to use FW 4.7.0
  
In case you need help with FW burning please approach our FAE Todd (cc here too).
 
After you have new results please reply if the performance numbers are OK now.
 
Thanks,
Tziporet
 
 -----Original Message-----
From: Galen M. Shipman [mailto:[EMAIL PROTECTED]]
Sent: Tuesday, July 26, 2005 5:55 PM
To: Tziporet Koren
Subject: Re: [openib-general] Shared Receive Queue

The following is from vstat.. 

 hca_id=InfiniHost0
        vendor_id=0x02C9
        vendor_part_id=0x6278
        hw_ver=0xA0
        fw_ver=0x400050003
        num_phys_ports=2
                port=1
                port_state=PORT_DOWN
                sm_lid=0x0000
                port_lid=0x0000
                port_lmc=0x00
                max_mtu=2048

                port=2
                port_state=PORT_ACTIVE
                sm_lid=0x0001
                port_lid=0x010a
                port_lmc=0x00
                max_mtu=2048


On Jul 26, 2005, at 8:46 AM, Tziporet Koren wrote:

which HW & FW are you using?

-----Original Message-----
From: Roland Dreier [mailto:[EMAIL PROTECTED]]
Sent: Tuesday, July 26, 2005 5:38 PM
To: Galen M. Shipman
Cc: [email protected]
Subject: Re: [openib-general] Shared Receive Queue


    Galen> We have software which uses the mellanox shared receive
    Galen> queue and we are seeing a substantial performance penalty
    Galen> (on the order of 4-5 usec on 0 byte messages). I don't know
    Galen> all the technical details but if you could keep this in
    Galen> mind when working on the shared receive queue perhaps the
    Galen> performance penalty could be much less in gen2.

Is this performance penalty in the software?  It seems much more
likely that the issue is in the hardware implementation of SRQ.

 - R.
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general

To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general



_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general

To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general

Reply via email to