On Apr 28, 2009, at 12:14 PM, David Anderson wrote: > At this point I'm interested in reports of bad scheduling > or work-fetch decisions, or bad debt calculation. > When these are all fixed we can address the efficiency > of the scheduling calculations (if it's an issue).
Sadly, in my opinion, this gets hard to do because of the needle in the haystack syndrome. Just as we finally acknowledged that there is an issue with checkpoints and the checkpoint interval with modern systems we need to come to grips with this issue as well. When you have events cascading at a rate of as many as 6 per minute the mass of data swamps investigation of issues. In that this is, theoretically, an easy patch, this would allow us to look at the deeper causes and more clearly expose them. I know John keeps reading this as I am suggesting that this is THE cure. It isn't ... but it is a most necessary first step. As a secondary effect, the rapid firing of triggers introduces its own level of instability. Another issue that obscures the underlying issues. I will also grant that it is entirely possible that the current system is wonderful for single and dual core systems. Let me gin up a pseudo code framework and try to demonstrate what I mean with that ... heck, nothing else seems to work ... _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
