>Are the balances you're doing on wakeup global or node-local?
The test is not done on NUMA systems.
IBM Linux Technology Center - Kernel Performance
"Martin J. Bligh"
<mbligh@...> To: Mala Anand/Austin/IBM@..., Erich Focht <efocht@...>, linux-kernel
Sent by: <linux-kernel@...>, LSE <lse-tech@...>
lse-tech-admin@... cc: Andi Kleen <ak@...>, torvalds@...
ceforge.net Subject: Re: [Lse-tech] Re: [patch] scheduler fix for 1cpu/node case
07/29/2003 09:29 AM
>>> If you want data supporting my assumptions: Ted Ts'o's talk at OLS
>>> shows the necessity to rebalance ASAP (even in try_to_wake_up).
>> If this is the patch I am thinking of, it was the (attached) one I sent
>> which did a light "push" rebalance at try_to_wake_up. Calling
>> at try_to_wake_up seems very heavy-weight. This patch only looks for an
>> cpu (within the same node) to wake up on before task activation, only if
>> task_rq(p)->nr_running is too long. So, yes, I do believe this can be
>> important, but I think it's only called for when we have an idle cpu.
> The patch that you sent to Rajan didn't yield any improvement on
> specjappserver so we did not include that in the ols paper. What
> is described in the ols paper is "calling load-balance" from
> try-to-wake-up. Both calling load-balance from try-to-wakeup and
> the "light push" rebalance at try_to_wake_up are already done in
> Andrea's 0(1) scheduler patch.
Are the balances you're doing on wakeup global or node-local?
This SF.Net email sponsored by: Free pre-built ASP.NET sites including
Data Reports, E-commerce, Portals, and Forums are available now.
Download today and enter to win an XBOX or Visual Studio .NET.
Lse-tech mailing list