ets:update_counter/3 taking a looooong time!

[Peter-Henry Mander]

Hi Gurus,

I'm puzzled.

I'm trying to build a stressor tool using the Megaco stack provided with 
Erlang/OTP R9B, and using fprof I seem to have isolated where the 
bottleneck is, except I don't believe it's in ets:update_counter/3.

The stressor creates a bunch of processes which run through a standard 
add modify and subtract Megaco message sequence. It runs wonderfully 
fast as long as there are no more than 16 or 17 processes started 
concurrently, each sending the message sequences. Below this threshold I 
manage to get 1500 calls setup per second, above I only get a pitiful 4 
to 6 call setups a second. If I increase the process count gradually to 
2000, I can start 16 new processes and still get a high rate, but any 
more than 16 at once kills the setup rate again.

I get the impression that some sort of staggered process startup is 
required, it's as if there's a process being starved somewhere.

The attached gives the output of fprof, and the last line indicates that 
all the time is spent waiting for ets:update_counter/3 to return a new 
transaction ID to megaco_config:incr_trans_id_counter/1. The functions 
above in the deleted lines creates a Megaco ActionRequest record.

What have I done wrong? I feel as if there's a solution staring me in 
the face, but I just don't see it!

Pete.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: stressor-analysis.txt
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20030428/ef244c34/attachment.txt>