Fun with threads#

Fri, 28 Feb 2003 03:17:54 +0000

Fun with threads. No, really. No sarcasm.

So after cleaning up the CVS mess left over from yesterday's merging exercise, I checked out a fresh copy of Araneida, built a new threaded SBCL, and hacked around a little.

1. When using apachebench for microbenchmarking your web server, the easiest way to get a really good requests/second figure is of course to use a really small file. I was seeing 90-130 requests/second at times, but each response is something like 240 bytes long, so, well, yeah.

2. When ab complains about failed requests due to "length", what it means is that some of the documents returned were of different length to the first response. Obviously if your test document is anything like this one, you can expect 91 alleged failed requests in the first test in a fresh image.

3. So, as you'd guessed, I'm using apachebench as a basic smoke test to flush out the most obvious threading issues in SBCL. So far I found
   • that my spinlock_get in C actually didn't (didn't spin or lock, in fact),
   • that there was some thread-unsafe code to allocate new buffers for file-streams,
   • that my db-sockets hack to allocate new sockaddrs in Lisp space really needed to be protected with without-gcing,
   • but that it wouldn't have made much difference if it had been, because without-gcing doesn't actually disable garbage collection any longer.

   Fixed the first three, kludged around the last - now it would work if atomic-incf were actually atomic instead of just a placeholder for incf.

   All told it's going pretty well right now, so there's bound to be some horrible problem cropping up soon.