Karel Gardas wrote:
Erik,
thank you once again for very valuable recommendation. I'm distributed systems
guy and plan to use the box for testing various messaging/rpc implementations
with CORBA on the start going thorough various SOAs and MOMs. Now, with CORBA,
MICO concretely (www.mico.org). Imagine this is just like an ordinary web
server accepting requests and sending back replies. The concurrency model in
use is thread per connection. i.e. one TCP connection serving one CORBA client.
Whole request/reply business is done in the thread together with all needed
memory allocation/deallocation. Side note: Yes, I've been already hit by
scalability issue with standard Solaris ST allocator so now I'm using umem by
default. Anyway, if you have a look at this I would imagine that such
application might scale very well as the allocated memory should/might be
placed on the CPU running thread connected RAM. Well, I don't know how Solaris
will handle the NIC/TCPIP stack here, but application alone might be q!
uit
e ok even running on 4x dual-core CPUs NUMA box, don't you think?
Thanks!
Karel
IIRC, the standard network stack runs on CPU0 (along with many other
kernel items), though it might need some tuning for your particular app.
Otherwise, your app definitely sounds like one which should scale well
across a 4- or 8-socket system. The memory allocator should definitely
prefer "local" memory when initially spawning the thread, and it sounds
like the thread's work is light enough that it should complete in a
single time-slice (that is, if it can do all it's work before the
scheduler decides to give another thread the CPU). You may have to look
at increasing the time-slice value (or, possibly the priority of the
app) to make sure that you get the thread done before it is forced to
sleep. If a thread is forced to sleep by the scheduler, well, then
you're in the problem that I described, where when it wakes back up, it
may not be assigned to the CPU where it's data is in local RAM.
--
Erik Trimble
Java System Support
Mailstop: usca22-123
Phone: x17195
Santa Clara, CA
_______________________________________________
opensolaris-discuss mailing list
opensolaris-discuss@opensolaris.org
