Re: Poor RAC inter-instance performance

Ivan Vasquez wrote:

> If I run the query in B first, it will perform well. Then if I go to A and
> run it, A will now be the slow instance. Bad performance can happen at
> either instance.
>
> Furthermore, when the query result is reasonably small (probably completely
> cached in memory), any subsequent runs at the badly performing instance will
> now run well.
>
> Under these conditions, even an unrestricted select * from table can be
> submitted and hang just as if it was processing a complex where condition.
>
> I tested the interconnect by scp'ing a 300+MB patch file: It transfers in
> matter of a few seconds (and I do see the link busy!)
>
> The machines have dual PIII 700Mhz. The queries are run while logged in
> directly into the node to discard network slowdowns. I patched them to
> 9.2.0.5 to discard any known bugs.
>
> We need to find a solution for this before we decide to implement RAC in
> production, as even if it is with better hardware, they would be i386
> machines as well.
>
> Thanks again.

My guess is that the problem is with the parameters in your sysctl.comp file. Most likely the system is choking on the TCP/IP stack. So look at dialing up the RMEM_MAX, WMEM_MAX, and some of the other hardware specific sendings.

I'd suggest running a statspack level 10 and a hang analyze. You can find information on hang analyze at metalink.

If none of this works post your sysctl.comp.

-- 
Daniel Morgan
http://www.outreach.washington.edu/ext/certificates/oad/oad_crs.asp
http://www.outreach.washington.edu/ext/certificates/aoa/aoa_crs.asp
damorgan_at_x.washington.edu
(replace 'x' with a 'u' to reply)