Victor Duchovni: > On Tue, Sep 07, 2010 at 09:07:54PM +0200, lst_ho...@kwsoft.de wrote: > > >> Single-core CPU limit. The system had 4 CPUs and the load peaked at ~25%. > >> The queue manager is single-threaded, and must do a fair amount of message > >> envelope processing. So the current design tops out at ~2-3k msgs/sec, > >> which is substantially faster than other constraints on real systems, so > >> the queue manager is not your bottleneck in real systems. > > > > Ok, so if one wants to really peak out it is more useful to have less > > cores, but faster ones given that I/O is able to keep up. > > Purely hypothetical discussion, no real MTA handles > 1k messages every > second. Yes, a single faster core improves the impractically high ceiling > on queue manager throughput when the disk subsystem is so fast that the > queue manager is CPU rather than I/O constrained. Today's practical queue > managers with queues on spinning disks will get bogged down in I/O first.
Large persistent-memory file systems are becoming available. Some emulate disks and therefore are constrained by disk protocol overhead, while others will run at native speed using file systems that are optimized for FLASH-like technologies. I wonder what the next bottleneck will be: network thoughput/latency, 1000s of mail delivering processes thrashing caches while switching context on many-core systems, or the queue manager. Wietse