Re: [Sbcl-devel] Binary release sb-thread policy

[Nikodemus S. <nik...@ra...>]

On Fri, Nov 7, 2008 at 1:18 AM, Justin Grant <jgr...@gm...> wrote:

> why would the same application code work just fine under
> heavy load on single core systems but become very unstable on multi-core
> systems ?

Timing issues.

Assume there is a non-thread-safe codepath that is relatively short
(say a couple of dozen instructions, no syscalls.)

To see the problem on a single core system you need to hit a context
switch in the middle of it, *and* then get the other thread to execute
the same codepath. On a multicore system your chances of a context
switch are multiplied (assuming same average time-slice size), and
additionally you can get two threads to really execute the piece in
parallel.

The *real* parallelism is the bit that makes the big difference, but
more context-switches / real-time-unit helps too.

Cheers,

 -- Nikodemus