64 processors? [LWN.net]

64 processors?

Posted Sep 21, 2007 22:41 UTC (Fri) by ms (subscriber, #41272)
In reply to: 64 processors? by roelofs
Parent article: What every programmer should know about memory, Part 1

I'm not totally sure what the author's referring to, but hyper threading is coming back. Of course, the Sun Niagara chip does 4 threads per core, rather than two, and IBM's Power series has been doing SMT (hyper threading by another name) for some time. http://anandtech.com/cpuchipsets/intel/showdoc.aspx?i=310... talks about 8 core Intel CPUs that will support hyper threading per core for 16 threads per CPU.

to post comments

64 processors?

Posted Sep 22, 2007 21:57 UTC (Sat) by deater (subscriber, #11746) [Link] (6 responses)

Please realize that Intel's "Hyperthreading" is not the same as "Multi-threading" as implemented by other vendors. Hyperthreading will often give you worse performance than simply using one core (ie, turning hyperthreading off).

Multi-threading, as in real SMT, will actually give you a performance improvement, but just not a linear one like adding an extra full core would. For example, extra hardware threads on the Niagara give about the same improvement as about 50% of a CPU.

I've done tests to verify this on our servers, and as a result I disable Hyperthreading on all of our compute nodes. Don't take my word for it though, run tests yourself.

64 processors?

Posted Sep 23, 2007 15:33 UTC (Sun) by alankila (guest, #47141) [Link] (4 responses)

In my experience, running almost any two perl tasks in parallel on a HT tends to take longer in wallclock time than running the two one after the other (keeping the other hyperthread idle). The difference is not large, but it tends to be a few percentage points slower with hyperthreading than without.

I pondered this for a while, and thought that it might be because each hyperthread in turn pushes the cached data of the other hyperthread off the cpu, resulting in expensive memory re-fetches which in this case actually negate the advantage of utilizing the idle CPU computation units.

Even so, I'd still keep HT enabled for web servers and like. This is because the small loss in throughput is likely to be repaid by a small decrease of system latency. Two processors, even two "half-speed" ones, can still do two tasks at the same time, and short computation bursts can well occur faster thanks to increased likelihood of there being an idle core immediately available for the task.

64 processors?

Posted Sep 24, 2007 15:47 UTC (Mon) by giraffedata (guest, #1954) [Link] (2 responses)

thought that it might be because each hyperthread in turn pushes the cached data of the other hyperthread off the cpu, resulting in expensive memory re-fetches which in this case actually negate the advantage of utilizing the idle CPU computation units.

It's roughly that effect, but this actual scenario doesn't happen, because a hyperthread timeslice lasts only until the next cache miss and there is more than one line of cache. The point of hyperthreading is to let Thread B run while Thread A is waiting to get stuff into cache.

However, you do have the simple problem of having only half the locality in the memory cache. If two threads are in greatly different contexts, each will get half the cache and half a cache might make it take significantly longer to get something done than a whole cache.

But I think the usual reason for hyperthreading slowdown is just the overhead of switching threads. If that exceeds the time you would sit idle waiting for memory access in the single thread case, the double thread has less throughput than the single thread.

What should have been said along with the 50% number above is that it is just for some highly average workload. The effect of hyperthreading has a great deal to do with how much memory access the programs do. A system that spends most of its time moving stuff around in memory will not benefit from hyperthreading. But a system that spends half its time accessing memory and the other half computing in registers and near caches might get close to double the throughput with 2 hyperthreads.

I usually turn off hyperthreading just to restore consistency. Otherwise, my speed varies almost randomly.

64 processors?

Posted Sep 25, 2007 10:05 UTC (Tue) by alankila (guest, #47141) [Link]

Thanks for a highly informative clarification.

Hyperthreading performance

Posted Oct 4, 2007 20:10 UTC (Thu) by anton (subscriber, #25547) [Link]

But I think the usual reason for hyperthreading slowdown is just the overhead of switching threads.

In SMT (and that includes hyperthreading), there is no thread switching overhead. The execution core just executes instructions from different contexts at the same time (but in different resources).

I don't know why the Pentium 4 variant of SMT performs as badly as it does; cache thrashing may contribute, but I don't think that this is the main reason. The main reasons are probably some obscure microarchitectural details, maybe the replay system, maybe something else.

64 processors?

Posted Sep 26, 2007 10:16 UTC (Wed) by epa (subscriber, #39769) [Link]

I don't think that is a fair comparison. You would need to compare running two perl tasks in parallel with HT versus running two in parallel without HT. That will tell you whether HT itself gives any performance improvement for parallel-perl workloads, and so may give a clue about whether to enable it on a production web server (which presumably is running lots of perl jobs at once).

Whether your two programs are best scheduled in parallel or serially is another question - as you found out, it may be faster to run one then the other. This does not necessarily imply that hyperthreading should be turned off.

64 processors?

Posted Sep 25, 2007 19:09 UTC (Tue) by csnook (guest, #36935) [Link]

Hyperthreading is generally not a win on workloads that are already well-optimized. If the application is making effective use of the cache, and is compiled to minimize pipeline bubbles, hyperthreading will just reduce your cache hit rate.

The problem that Hyperthreading solves quite well is workloads that cannot be well-optimized because they react to unpredictable external events, as is the case with web servers. It's also beneficial on desktop systems where having only one CPU can harm interactivity while background tasks are running, but in the age of multicore CPUs that won't be much of an issue.

As a rule of thumb, the more performance tuning you're doing on a system, the more likely it is you'll want to disable Hyperthreading. HPC nodes and centralized database servers definitely suffer, while build systems and static web servers definitely benefit.