linux-kernel-digest V1 #2975
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 22 lines · David S. Miller
From: [email protected] (Alan Cox)
Date: Fri, 11 Dec 1998 16:07:50 +0000 (GMT)One problem with this is that most vendors of kernels ship large complex OS's that will reliably have blasted the last task out of the caches by the time they lumber back around to running it again.
Going to main memory can be at such a high cost that going there at all is to be avoided at all costs. The Challenge series SGI SMP boxes had a huge main memory latency, so they worked like mad on the cache warming heuristics in the IRIX scheduler (so much so that their scheduler is overly complex, and this complexity only got worse when the NUMA support went in).
Point is, try but not too hard.
Alpha and Digital Unix took a slightly different approach. I believe from a paper I read (before I get bombed with emails, no I have no idea where this paper was nor the title) on their kernel SMP hacks that it was:
Quoted 7 lines
1) Per-cpu run queues
2) Fork added new task to parents processor run queue
3) Exec pushed the task to another cpu's run queue
4) Tasks stayed on the same queue normally
5) CPU idleness or overload caused migration of tasks to other cpu run queues if the process load was not equally distributed at such time
One of the same group's papers did some work with this, too. Their only interesting result was, IIRC, that the synchronization was more important than the cache. They found that a data structure that was owned by each CPU had a slight edge. The competition for a single run queue *did* impose a noticable barrier.
Quoted 2 lines
I get the general effect of 2 and 3 and a little bit of 4 on sparc64/SMP with this tiny hack I put into the idler loop:
I am a bit sheepish to say it, but can you show any rigor to your results? I ask because my understanding with this level of optimization is that very little of it has shown to be always true. Indeed, intuition does not always grok silicon.
Quoted 11 lines
if (current->need_resched != 0 || ((p = init_task.next_run) != NULL && (p->processor == smp_processor_id() || (p->tss.flags & SPARC_FLAG_NEWCHILD) != 0))) schedule();It's a nice cheap heuristic. And actually I added it so that idlers didn't bang into the scheduler (grabbing locks, blowing dirty cache lines across the bus ping pong style, etc.) over and over when no new work was available. This nasty behavior is how ix86 SMP behaves currently until something equivalent is added to it's idler loop.
Hmm.
Quoted 3 lines
Later,
David S. Miller
[email protected]
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
Please read the FAQ at http://www.tux.org/lkml/