On 09/12/2008 5:28 AM, Henrik Krohns wrote: > On Tue, Dec 09, 2008 at 10:00:03AM +0000, Justin Mason wrote: >> Daryl C. W. O'Shea writes: >>> The long standing bottleneck (for net-enabled mass-checks) in my >>> otherwise fast mass-check cluster is Bind. It seems that it simply >>> cannot handle the load of a dozen or so cores worth of mass-check >>> processes. 475 kmsgs/hr non-net versus 70 kmsgs/hr net-enabled is >>> unacceptable. >>> >>> What are people using for a high capacity DNS recursive resolver... >>> which hopefully includes cache capabilities? >> well, I don't have a mass-check cluster ;) >> >> What about multiple binds, one per machine? > > I think multiple nameservers make it even worse. I don't see why a single > decend nameserver wouldn't work for a _tiny_ load like this. Bind even works > on multiple cores.
I agree, Bind can perform way better than it is. I've likely borked something along the way... actually thinking about it I think a distro update may have happened around the same time things went downhill. > We are missing the crucial details.. Yeah, wasn't really looking to troubleshoot at this point, just looking to see if anyone was using something else with good results so that I could compare both. > Are all queries to local mirrors? No, but it's not delays in waiting for responses, it's Bind chewing CPU like crazy and becoming slow to resolve anything at all. > Are concurrent processes raised accordingly when using net-enabled? You have > to take in account the DNS delay that's slowing you down. Not to mention > Net::DNS is a resource pig. It's not an issue with the clients. It's definitely the recursive server. There's probably something obviously wrong with its config. I only looked at it quick to confirm that its not swap thrashing and its eating CPU. Anyway... thought I'd see what else people were using. Daryl
