[PRL] Re: [Fwd: CSAIL Event Calendar] TALK:7-14-04 Prof. Mor Harchol -Balter From CMU

Rajmohan Rajaraman rraj at ccs.neu.edu
Wed Jul 21 18:53:41 EDT 2004


On Wed, 14 Jul 2004, Kenichi Asai wrote:

> Hi there,
>
> I attended Mor's seminar.  Her talk was impressive.  She showed that
> SRPT (Shortest Remaining Processing Time) policy is superior than the
> simple PS (Processor Sharing, or the standard time sharing) in all
> cases regardless of the processing time of each job.  It appears that
> she obtained this conclusion after studying and experimenting almost
> all the conceivable cases.

This work is somewhat old -- I think the original papers of Mor's appeared
more than two years ago.  The (provable) superiority of SRPT over other
policies, with regard to metrics such as average response time and average
slowdown, is also documented in other studies.

The above work on SRPT focuses on a single resource (e.g., CPU).  From
what I understand, some of the more recent work of Mor's group is to
conduct a performance study taking into account all the resources that a
server may need -- CPU, I/O, network interface, etc., -- and their
interactions.

>
> It was also interesting to hear that we can apply the lesson to our
> daily life.  When a professor has too many students, he should better
> finish a smaller task first rather than allotting equal amount of time
> to each student.  This would reduce the mean response time.  I feel
> this actually coincides with the experience.
>
> She talked most about static documents, but she says her current focus
> is on dynamic documents.  I thought that if you want to work in this
> area, she is the right person to contact.

The SRPT-style approach applies to static documents only, since it is
assumed that the processing time (as well as the remaining processing
time) of a job can be computed at any time.  Perhaps, the "dynamic
documents" work that Mor is referring to is their ICDE'04 paper on
priority-based scheduling of transactional web applications (available
from her webpage).

Rajmohan.



More information about the PRL mailing list