Re: Multiple instances of Splus on dual CPU

["tomas mann" <mann_tomas@hotmail.com>]

It turned out that the main culprit is the hyperthreading (at least for Win 
2000 server) - when I turned it off (on a dual CPU machine), I managed to 
run in parallel 2 identical S+ processes for roughly the same tyme (OK, 
maybe 1.2-1.3 times slower) as a single process on an identical 1 CPU 
machine.
A summary of the responses is enclosed. TM

Nels Tomlinson:
> I wonder if you'd find that turning off hyperthreading would cause both
> processes to complete in five minutes?
>
> Hyperthreading is a clever trick to prevent CPU resources from sitting 
> idle,
> in some circumstances.  When you have two processes which don't require the
> same CPU resources (e.g., Splus and Word), hyperthreading lets both of them
> execute almost simultaneously.  Splus can hog the FPU without entirely
> blocking Word.  When you have two processes which require the same 
> resources
> (e.g., two instances of Splus, which both heavily tax the FPU), they will
> contend for those resources, and at best won't slow each other down.
>
> Remember, you really have two CPU's, and hyperthreading doesn't change 
> that.
> At best, it can help you use those two more efficiently, and at worst 
> you'll
> want to turn it off for some jobs.  This could be one of them.
>
> Another possibility would be that the process is I/O bound.  Modern CPU's
> are usually I/O bound, for most tasks, since MHz has been hyped for sales
> purposes, but bandwidth hasn't.  If getting stuff in and out of RAM (or
> worse yet, on and off the disk)is the bottleneck, then hyperthreading
> doesn't matter.  I would expect that situation to result in two tasks 
> taking
> twice as long.  Since you say that the CPU's aren't idling, that doesn't
> sound like the case here, but I'm not familiar with the details of the
> Windows resource monitors.  What are the CPU's busy at?  Can you look at 
> the
> memory and disk utilization patterns?
>
> I guess that you'll have to identify the bottleneck, and see whether that
> second CPU is opening it up or making it tighter.
>
> Some interesting information would be whether your machine uses SDR, DDR or
> RDRAM (they have different latency and bandwidth characteristics), what the
> bus speed is and whether your code loops tightly on a smidgen of data, or 
> is
> pumping gigabytes through the CPU's.  Is the computer swapping with one
> process?  With two?  That sort of info might help in guessing about where
> the bottleneck lies.
>
> One final thought: could it be that the two instances of Splus are
> contending for one copy of some other resource, such as a dll, or standard
> output?  Again, I'm not very familiar with Windows, so I don't know if
> that's a realistic suggestion.
>
> If you get any hard information about what's happening, I'm sure that I'm
> not the only one who'd like to hear it.

John James, Insightful
(This is an informal reply and does not represent the view of Insightful)
One suggestion - and I don't know of a computational way to implement it 
yet,
although it's on my "to do" list - is to use the task manager to assign each
S-PLUS process to each CPU.
What happens in 2 and 4 cpu tasks is that the OS tries to load balance and 
swaps
memory. Therefore its common that S+ runs slower on 2 cpu machines cf. 1 cpu
machines.

Vadim Kutsyy:
I would say that there is something wrong with your system setup. I
would check use of SWAP file. I often run multiple processes of Splus on
dual Xeon (Suse Linux) and Dual Pentium 4 (Win2K).  Both works as
expected. :
f<-function(n) for(i in 1:n) y=sqrt(i)
Single procces:
> sys.time(f(1000000))
[1] 105.109 107.058
Dual process:
> sys.time(f(1000000))
[1] 106.297 109.621
> sys.time(f(1000000))
[1] 106.046 109.136

_________________________________________________________________
Find a cheaper internet access deal - choose one to suit you. 
http://www.msn.co.uk/internetaccess