Linux MultiCPU performance

On Mon, 30 Jan 2006 14:29:03 +0200, Scott Hazelhurst

How many actual processors do you have? If it's two hyperthreaded Xeons then you expect exactly the behavior that you are seeing. If a P4 core could run a single process perfectly efficiently then running two threads on the same chip would yield 0% improvement in throughput and each individual thread would take twice as long to run as it would if you ran it by itself. However the P4 core architecture is dreadfully inefficient because the pipelines are way to long. The long pipelines mean that each incorrectly predicted branch cause the processor to throw away a huge number of instructions that were already in the pipe. When you run two threads in a hyperthreaded manner you've effectively cut the pipeline depth in half so the penalty for wrong branches is much less. The net effect is that the throughput when running hyperthreaded is about 30% greater (at best) then running a single thread. On top of that dual Xeon systems share a common main memory system so the gain from the second processor is less than 2x. With hyperthreading turned on you should expect a dual Xeon system to have the thoughput of no more than 2.6 processors, it could even be less. Under no circumstances would it ever look like 4 processors.