Re: Load balance strategy

From: Gengbin Zheng (gzheng_at_ks.uiuc.edu)
Date: Fri Mar 19 2004 - 01:41:44 CST

On Thu, 18 Mar 2004, Alex V. wrote:

> Thanks for all your advises!!!))
>
> I've already tried to reduce firstLdbStep parameter and gained
> performance of my calculations. I really needed that while simulating
> *huge* molecular system. But I think it is more useful on heterogeneous
> clusters when more powerful nodes waste there resources waiting
> for slower PCs (at startup phase).
>

That is great! Since the load balancing strategy is adaptive, it will work
for *heterogeneous* clusters as well. But I hope by "heterogeneous" you
meant clusters of same kind of machine but with different CPU speeds.
NAMD and Charm++ do not support real heterogeneous cluster yet such as a
cluster with mixed PC Linux and SUN workstations.

> By the way, will *Charm++* become an international standard for
> parallel processing? Are you planing to provide integration with CORBA?
>

It is a standard, ;-) one day.

There is no plan to integrate with CORBA though. I don't see they are
relevant. CORBA is a middleware for writting client and server type of
applications. It does not target on large parallel computations, it may
focus more on providing distributed services.

Gengbin

> Alex
>
> Gengbin Zheng wrote:
>
> >NAMD load balancing capability is based on Charm++ (not charmm :).
> >It is a measurement-based load balancing. That is the load balancing
> >decision is made adaptively during run-time based on load statistics
> >instead of the static mapping. You may have seen the startup display:
> >
> >Info: Entering startup phase 6 with 7884 kB of memory in use.
> >Measuring processor speeds... Done.
> >Info: Entering startup phase 7 with 7885 kB of memory in use.
> >
> >There is a measurement of CPU speed at the startup. The cpu speed should
> >be already considered as a factor during load balancing decisions.
> >Due to the measurement based strategy in nature, it does need several
> >steps of simulation running before it can actually have enough
> >information for load balancing decision making.
> >One can use "firstLdbStep" to control the first time to start the
> >first load balancing and "ldbPeriod" to control the loadbalaning period.
> >
> >Gengbin
> >
> >
> >
> >>I think charmm can have a "speed" directive that can do that.
> >>
> >>Ciao
> >>
> >>
> >>
> >
> >
> >
> >
> >
> >
>

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:38:31 CST