4 ms·
If it is an I/O heavy workload (and it is), then his statement "All 101 threads run 99.9% of the time and we get an almost 100× speed-up." cannot be true. The p
by gvb 10y ago
If it is an I/O heavy workload (and it is), then his statement "All 101 threads run 99.9% of the time and we get an almost 100× speed-up." cannot be true. The processing portion probably runs n times faster, where n = the number of CPU cores, but his I/O thread is still single-threaded ("[w]e cannot do writing to disk from several threads for the following reasons...") and obviously cannot be running 100x faster.
- dllthomas 10y agoWell, it's possible that the IO is dispatched from a single thread but can be handled in parallel by the OS. If the single thread doesn't need to block, it needn't be the bottleneck. Still, even if they are somehow seeing 100x speedup, it won't be because 100 threads are near constantly busy - without that many cores (after hyperthreading, maybe, with caveats) it is simply not going to be the case that all threads are constantly busy.