Why Can't 2 Projects Be Run At The Same Moment

As to why only one process per processor, that has to do with the large overhead for context switches. Most of the science applications have large data sets and already fill the L1 and L2 caches to capacity. Running several science applications on the same processor would increase the number of cache misses, and reduce the overall amount of work that can be done.

> Thank You John,
> That is exactly what I was after. I had remembered someone, it might have
> been you posting on the Seti Boards quite a while back and I could not find
> the post. I remembered because I had made a mistake when I bought my trhird
> box because I bought a Celeron Processor instead of a more effect processor
> with a higher L1 and L2 cache. I am just looking to maximize the efficieny of
> my boxes!
> So thanks for the help.
>
> Warm Regards,

Rocky,

What John said ... :)

Most people do not realize that a "context switch" is one of the more expensive operations in that you have to save the entier processor "context" meaning all of the registers, internal pipelines (or run them dry), where you are in memory, etc.

Then you load the other context and start it up. This is one of the reasons that Microsoft changed NT to move the device drivers into the "Ring 0" portion of the Intel processors ... it makes the system faster... unstable ... but faster as you don't have to do a context switch to change from OS space to Application space ...

This is another reason that "threads" are used. The complete term is "Thread of Execution" and you can have multiple threads operating within a process. Since the Process context remains the same, changing the thread of execution within a process is "cheaper" ...

With the multi-threaded processors, they can execute multiple threads maintaining seperate contexts for each thread ...

However, John caught the essence with the cache being even more critical in this, ahem, context, with the higher miss rates causing a serious performance penalty if you try to run multiple processes/threads on a single threaded processor core.

In SETI@Home Classic, people would brag about "forcing" the processor to run multiple instances of the program on single threaded processors to have 10 work units in process at once. Even when told they were wasting their time, and slowing their performance ... well ... some people are slow to switch their context ...

In my case, I have 6 computers, one a true dual processor, and one single threaded processor (2.8 GHz Intel) and the other 4 are all HT capable and do run two work units from one or two projects at the same time. It can be two work units from one project or one each from two projects. I am probably losing a little performance from cache "thrashing" but, throughput is good, and another Macintosh is in the plan for this year ...

With great luck it will be faster than 2.5 GHz, dual processor, dual core ... giving me 4 "logical" CPUs.

For those taking notes ... some of this is discussed in the Glossary under logical and Physical CPUs ... and when I get my lecture notes on-line you can read my take on history ... and how we have made the same mistake for every[/b] new class of computers from EINIAC on to PalmPilots ...

Oh, and I probably should say something about this in the FAQ ...