RE: Infiniband for RAC 10g on Linux

interconnect latency is not the problem. Once a node goes processor saturated,
you can get an skgxp transfer there
in sub nanosecond time, but the process
that needs the message is only runable.

what is the time quantum? 10ms, so
the reciever will eventuall get on a CPU to catch that really fast message

>>>-----Original Message-----
>>>From: oracle-l-bounce_at_freelists.org
>>>[mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Keith Moore
>>>Sent: Friday, September 09, 2005 1:19 PM
>>>To: Matthew Zito
>>>Cc: kmoore_at_zephyrus.com; Oracle-L Freelists
>>>Subject: Re: Infiniband for RAC 10g on Linux
>>>
>>>So, do you know the latency for IPoIB? Is it less than GB
>>>Ethernet? Or about the same?
>>>
>>>Thanks.
>>>
>>>Keith
>>>
>>>>
>>>> InfiniBand is fairly cheap. However, true IB for RAC
>>>(uDAPL) is not
>>>> supported in 10gR2. As far as we've been able to figure out, the
>>>> InfiniBand support is for SDP and is undocumented (that's for
>>>> SQL*Net) and for interconnect traffic is only IPoIB -
>>>standard IP run
>>>> over InfiniBand. Oracle has dropped plans to support uDAPL -
>>>> apparently deciding to try to push their own proprietary
>>>RDMA protocol.
>>>>
>>>> Matt
>>>>
>>>> On Sep 9, 2005, at 3:40 PM, Keith Moore wrote:
>>>>
>>>>> We are performance testing an Oracle 10g RAC setup with
>>>two RedHat
>>>>> Linux servers. Currently we are using Ethernet for the
>>>interconnect.
>>>>>
>>>>> Has anyone used Infiniband?
>>>>>
>>>>> How well did it work and how much can we expect to pay?
>>>>>
>>>>> Thanks
>>>>> Keith
>>>>>
>>>>> --
>>>>> http://www.freelists.org/webpage/oracle-l
>>>>>
>>>>
>>>>
>>>
>>>
>>>--
>>>http://www.freelists.org/webpage/oracle-l
>>>

--
http://www.freelists.org/webpage/oracle-l