On Tue, 8 May 2012, Erik Trimble wrote:
Roy hits the biggest nail on the head, here. HT should work just fine (i.e.
give you almost 2x CPU performance) with two major issues:
It seems to hardly ever do that for me. My impression is that HT is
about improving latency rather than to add real CPU potential. The
hyper-thread does not need to have its context reloaded so it is ready
to go again quicker once whatever resource which was blocking it is
releasd.
(1) anything where you're doing a large amount of memory transfers (i.e. your
workload is mostly operating on fresh data, rather than the results of prior
calculations), the lower amount of cache per thread really hurts, especially
if the two threads are huge cache hogs - e.g. doing large matrix
calculations. This ends up cache thrashing, and both threads suffer
horribly.
(2) lots of FP work, since there's still only one FP execution unit being
shared between the two threads.
The above describes a whole lot of software. Software not matching
the above description is likely not CPU-bound.
Real world experience/metrics would be beneficial here.
The plot at
"http://www.simplesystems.org/users/bfriesen/zfs-discuss/thread-scaling.png";
represents a typical experience I have seen with hyper-threading.
This is for a 4 socket Xeon system with 16 cores and 32 threads. I
have yet to see hyper-threading cause enough harm to want to disable
it.
The system is running Linux kernel 2.6.32. This OS selected to use
the real cores first before the hyper-threads. The Ubuntu Linux
system I have here alternates between a real core and a hyper-thread.
If one looks at the posted benchmark results on the SPEC site, one
will see that vendors enable or disable hyper-threading (sometimes a
vendor posts results on same hardware for both) and the difference in
the results is quite marginal.
Of course, the difference between the -1220 and -1230 is under $25, so I'd
probably chance it with the -1230, as you can shut off HT if it doesn't work
out.
If the price difference is this small, then it is best to retain the
option.
Bob
--
Bob Friesenhahn
[email protected], http://www.simplesystems.org/users/bfriesen/
GraphicsMagick Maintainer, http://www.GraphicsMagick.org/
-------------------------------------------
illumos-discuss
Archives: https://www.listbox.com/member/archive/182180/=now
RSS Feed: https://www.listbox.com/member/archive/rss/182180/21175430-2e6923be
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21175430&id_secret=21175430-6a77cda4
Powered by Listbox: http://www.listbox.com
![http://www.simplesystems.org/users/bfriesen/zfs-discuss/thread-scaling.png"](http://www.simplesystems.org/users/bfriesen/zfs-discuss/thread-scaling.png"){kind=link}