pcp
[Top] [All Lists]

Re: PCP developers meeting notes

To: Mustafa Sezgin <msezgin@xxxxxxxxxxxxxxxx>
Subject: Re: PCP developers meeting notes
From: Nathan Scott <nathans@xxxxxxxxxx>
Date: Mon, 27 Jun 2011 15:17:07 +1000 (EST)
Cc: Mark Goodwin <mgoodwin@xxxxxxxxxx>, Max Matveev <makc@xxxxxxxxx>, David Chatterton <dchatterton@xxxxxxxxxx>, Gokul Krishnan <gkrishnan@xxxxxxxxxx>, Ken McDonell <kenj@xxxxxxxxxxxxxxxx>, Paul Smith <psmith@xxxxxxxxxx>, "Josef 'Jeff' Sipek" <jeffpc@xxxxxxxxxxxxxx>, pcp <pcp@xxxxxxxxxxx>
In-reply-to: <BANLkTin2L8g-tyk7cpmzOu+LM9YUYGA6OA@xxxxxxxxxxxxxx>
----- Original Message -----
> Can I get some more info with regards to the issues with Parfait? What
> are the scalability/critical issues that were seen?

There were a few.  The first class was self-inflicted (and IIRC inherited
from Parfaits precursor) - the use of a single writeout queue with single
lock - it was always intended (and the C API implements this) that each
metric/value is written independently of any others (within the mmapped
file).

The second class of problem surrounds scaling issues within the JVM - in
particular, there is use of the ThreadLock mutex around calls from one
thread accessing threadInfo structure for another thread (problem doesn't
exist for a thread accessing its own threadInfo, that has a fast path with
no lock).  Once many threads are doing this, inter-thread contention bites
badly and the application can no longer scale to even modest numbers of
CPUs.

So, avoid those sorts of things, please.  :)

cheers.

-- 
Nathan

<Prev in Thread] Current Thread [Next in Thread>