SP Crashing

[Cantor, Scott]

On 9/19/19, 1:55 PM, "Cantor, Scott" <cantor.2 at osu.edu> wrote:

> Just avoiding prefork on Unix. The stack size should be much less than the default is, but it doesn't really matter as
> long as prefork isn't used. The thread count shouldn't ever get very large as long as the Apache thread count is
> reasonable.

I should perhaps not be so blasé about this, since I don't have any history with Windows usage and how it might behave under real world high loads. The default stack is, I think 1M, which is still nuts, and can certainly be drastically lowered and is a good simple step to take (it's documented how to lower that). If load spikes are for some reason not getting in and out of shibd fast enough to keep the worker thread count inside it from spiking, that definitely breaks on Linux and I suppose could break on Windows. But the size is bigger on Linux, so that's one reason the spikes are more fatal there.

But all my experience is that shibd is fast enough doing its work that as long as there aren't 100 processes connecting to it, the thread size stays manageable. Each process pools client sockets to connect, but with 100 web processes, you get a minimum of 100 threads in shibd.

-- Scott