[Haizea] Final CPU Utilization

Borja Sotomayor borja at borjanet.com
Wed Jun 16 18:47:52 CDT 2010


Hi,

> So in analyzing full raw data, we should do something like that you did
> in haizea_swf2lwf class to remove ramp-up/ramp-down effects, Am I right?

Yes, that's right.

> Some more questions about that code in haizea_swf2lwf:
> 1. By introducing no_ramp_cutoff, we take care of the beginning 5% of
> workload? and also the end of workload?
> no_ramp_cutoff = from_time + ((to_time - from_time) * 0.05)

This eliminates the measurements during the first 5% of the experiment
(which is generally enough to remove ramp-up effects; there is a paper
about this, but I can't find it right now). At the end of the
experiment, instead of removing some percentage of the measurements,
we simply disregard measurements after the last lease is submitted.

> 2. In utilization_no_ramp calculation, we shouldn't consider if there is no
> waiting time for a job, then reduce interval. And why we are using this
> technique? we want to put jobs in the normal condition of a workload that is
> in the middle i.e. neither the beginning nor the end.
> if wait_time != -1 and submit_time + run_time >= no_ramp_cutoff:
>                     start_in_interval = max(no_ramp_cutoff, submit_time)
>                     end_in_interval = min(to_time, submit_time + run_time)
>                     time_in_interval = end_in_interval - start_in_interval
>                     utilization_no_ramp += time_in_interval * num_processors

Since this utilization is computed when converting a SWF trace to an
LWF trace, we don't have the benefit of knowing when the lease would
actually be scheduled; so, the above simply figures out what the
utilization would be in the best-case scenario that every lease
started at its earliest possible start time (which, for the purposes
of computing the utilization, requires truncating the duration of the
lease if some part falls outside the interval where we're computing
the utilization).

> Is it possible for me to access R infrastructure now?

I will contact you off-list about this.

Cheers!
-- 
Borja Sotomayor
PhD Candidate in Computer Science, University of Chicago
http://people.cs.uchicago.edu/~borja/


More information about the Haizea mailing list