Sun faces up to the 64 thread question with T2+

Single thread performance

The Achilles heel of the Coolthread series of chips is the very poor single thread performance. It's fine if you have an application that can multi-thread, and where the response time doesn't depend on how fast you execute a single instruction stream. Unfortunately there's a lot of those sort of applications out there that are affected. Things get worse on the Coolthreads processors once you start winding up the throughput - those 8 hardware threads on a Niagara 2 core contend for just two integer and one FP processing unit. As utilisation winds up then the threads in execution slow down further due to this contention so your hardware thread might start looking more like a 700MHz (or slower) processor than a 1.4GHz one. it's not as bad as it might appear, as the core can make use of processing resources when a thread is stalled for something like a memory access, but it can have a very visible effect. it's best to view these hardware threads as virtual CPUs in the same way that guest machines under VMWare can be slowed due to contention.

A single or dual socket Sparc64 won't pirate sales from Coolthreads - it will fill a gaping hole in SUN's line up for SPARC servers; an entry level server with half way decent single thread performance. The Niagara series will always appeal where you need SPARC and want the best throughput per watt (the SPARC 64 is anything but power efficient), but the coolthreads architecture is not the answer to everything.

Nb. this whole things about hardware thread support is something that affects many modern processorrs SPARC64, Itanium and Power all have support for a coarse-threading model (and Intel used to have HyperThreads). They all trade a bit more throughput for much reduced single thread speed at high utilisation levels (and distinctly non-linear relationships between reported CPU utilisation and throughput).


Back to the forum


Biting the hand that feeds IT © 1998–2017