8/9 Linux question. We have a simple server that recieves TCP/IP
connections concurrently with a threadpool, creating new threads
as necessary. It's showing a weird performance quirk where, if
you increase the number of concurrent connections, the connection
time increases slowly, from .1s. At 16 connections it's
about .2s. However, the 17th connection takes 1.2s, a large
jump. The connection times continue increasing slowly from
there, although there are little jumps at 48 and similar
multiples of 16. slowly. Is there some magic kernel number 16,
above with establishing a TCP/IP connection takes a long time?
\_ Stupid question, does your threadpool have a max number of
threads?
\_ Yes, but it's 1024, which is actually higher than the kernel
seems to be able to generate.
\_ Java threads? pthreads?
\_ pthreads, it's all C or C++ code on chaos Linux,
(although we were able to duplicate the problem on
RedHat,)
\- i dunno what the linux equiv of tcp_conn_hash_size
is, but i'd personally be interested if changing
bumping that up changes the behavior.
is, but i'd personally be interested if bumping that
up changes the behavior.
\_ No. -proud American |