Charm++ parallel programming system

["Galvez Garcia, Juan Jose" <jjgalvez AT illinois.edu>]

Hi Anna,

 

Centralized load balancers (that is, those that inherit from CentralLB) work by sending statistics to a central PE (typically PE 0), who then runs a load balancing strategy to decide where to assign each chare, and subsequently sends the migration decisions to every PE. Since centralized load balancers all follow this pattern, and have the same inputs and outputs, that is why you can call the work() function of other centralized load balancers at PE 0.

 

DistributedLB however works very differently. The information is not gathered in a central location, and the assignment of chares to processors is calculated in a distributed fashion. So, you can’t invoke DistributedLB from a centralized load balancer. You could, however, probably tell Charm++ that you want to run DistributedLB after your own load balancer has finished. By this I mean that you have to wait for centralized load balancing to send stats to PE 0, calculate a solution, send migration decisions, and move the chares. And then you could run DistributedLB. But I am not sure this is what you want.

 

If you have more questions, don’t hesitate to ask.

 

-Juan

 

From: Anna Victoria Oikawa
Sent: Friday, February 8, 2019 2:51 PM
To: charm AT lists.cs.illinois.edu
Subject: [charm] Invoking DistributedLB from a CentralizedLB

 

Hello,

I'm developing a centralized Load Balancer that should be able to invoke other balancers.
Currently, I invoke centralized strategies following the examples in AdaptiveLB (calling their respective work() functions).

How can I invoke the strategy() function from DistributedLB?