dev/multi-process.md

new file mode 100644@@ -0,0 +1,30 @@+From Rob Hoelz:++I did some additional testing of ack using multiple workers -+starting with a ondemand forking model, and then moving to a pre-fork+model. The former didn't provide too much of a benefit, and the+latter was somewhat complicated. At first I thought maybe I could+have a shared work queue pipe, but I realized that each child process+would probably buffer inputs, so the work would get divied up oddly.+I then had the thought to maintain a list of available children for+work - children would send a SIGUSR1 to their parent when done, and+the parent would re-register the corresponding child in the+availability list. A good idea in theory, but it turns out that+even though `si_pid` is documented as occurring in `siginfo_t`, it+either isn't available on Linux, or the Perl POSIX module doesn't+pass it through. I ended up opting for a simple list of children+and iterating through them in a round-robin fashion. A possible+improvement could be random selection, perhaps weighted by the sizes+of the files that each child has handled thus far.++The multi-process model didn't show large benefits - although my+first experiment didn't actually perform any work. When I added+experiments to actually test against regexes, that's when I saw+about a 200% speedup running with 8 jobs (8 being the number of+cores in my machine). I tried 16 (figuring one job could be waiting+on I/O while the other used the CPU), but actually saw a slowdown+as a result. This might be different on HDDs, though.++Another big improvement was to use sysread with a 128K buffer (which+is the default cat uses) rather than Perl's diamond operator - this+combined with the multi-process stuff lead to a 400% speedup.