From: Rick L. <ric...@us...> - 2003-04-18 18:07:40
|
Dave asked: Rick, have you been able to examine the SPECjbb200(tm)* problem and figure out whether it is really a balancing problem? To answer your question directly, no, but I think as Andrew demonstrated much of the work has alredy been done! :) Thanks Andrew. JBB really shouldn't behave that way, but IMO, the scheduler should do its best should that situation arise. A change to wake_up helps this, since it's tied much more closely to the rate of task activation. That's why the wake_up modification works, and makes the NUMA config'd kernel run just as fast as the non-NUMA config'd kernel Here's my jbb-specific concern. From what I understand of jbb, once the benchmark gets underway, there's a thread for each warehouse and it puts its shoulder down and goes to work pretty intensively. This behavior, though suggests it's going in fits and starts. Do we know which is correct? That is, does this focus on NUMA scheduling take the attention away from some other aspect of the Specjbb benchmark which is just as important? Is there some sort of resource contention such that it stops and starts so often? What is it stopping for? Rick |