Re: [patch] CFS scheduler, -v8

[Ting Yang <tingy@cs.umass.edu>]

Li, Tong N wrote:
> Thanks for the excellent explanation. I think EEVDF and many algs alike
> assume global ordering of all tasks in the system (based on virtual
> time), whereas CFS does so locally on each processor and relies on load
> balancing to achieve fairness across processors. It'd achieve strong
> fairness locally, but I'm not sure about its global fairness properties
> in an MP environment. If ideally the total load weight on each processor
> is always the same, then local fairness would imply global fairness, but
> this is a bin packing problem and is intractable ...
First, I am not assuming a global ordering of all tasks. As the current 
implementation, EEVDF should maintain virtual time locally for each CPU. 
EEVDF is a proportional  time share scheduler, therefore the relative 
weight and actual cpu share for each task varies when tasks join and 
leave. There will be not bin-pack problem for such systems.

I understand that bin-pack problem does exist in Real-time world. 
Suppose in a system has 2 cpus,  there a 3 tasks, all of which needs to 
finish 30ms work within a window of 50ms. Any 2 of them stay together 
will exceeds the bandwidth of one cpu. There is a bin-pack problem, 
unless the system has to be clever enough to break one of them down into 
2 requests of 15ms/25ms, and execute them on different cpus at different 
time without overlap, which is quite difficult :-)

In the proportional world, weights and cpu share are scale to fit with 
the bandwidth of a cpu. Therefore putting 2 of them on one cpu is fine, 
and the fairness for each cpu is preserved. On the other hand, moving 
one task back and forth among 2 cpus do give better throughput and 
better global fairness. I have not dig into the load balancing 
algorithms of SMP yet, so I leave it aside for now, first thing first :-)

Thanks !

Ting
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/