[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Computer for Trade Station



PureBytes Links

Trading Reference Links

On Thu, 9 Dec 1999 22:22:44 EST, you wrote:

<snip>
>>  Someone mentioned they've been reading up on this and get only 50%
>>  performance increase on 2nd CPU - this is NOT true.  (Hope that didnt come
>>  across wrong).  But for reasons as explained earlier you should see a 
>LINEAR
>>  performance increase as CPUs are added.
>
>I've never seen a real benchmark that supports this statement in SMP 
>(symmetrical multiprocessing) configuration, even on multi-million dollar 
>Alphas. I suppose it's possible to write a synthetic program that would come 
>close, but I've never heard of it in the real world. MPP (massively parallel 
>processing) is a different story, but you might need a couple million bucks 
>for a nice Pyramid or something like that.

I was going to say the same thing.  I've never seen a linear increase
either, even from benchmark.  There is always a performance penalty
(Below 100% of added CPU capacity)  for adding extra CPUs.  

Besides, in most systems today, the CPU is standing still waiting on
the slow [expletive] whatever (Memory, HD, Floppy, Video, ...).  

However, a recent change in Tradestation has dramatically increased
the CPU requirement for the software.  I've noticed that the problem
seems to be in the moving average functions.  If these are removed,
charts come up at the usual pace. (no, I haven't gone in and studied
this problem closely, just removed a indicator or two and retested)

As a programmer, I know that even a small change to an algorithm can
make a huge performance difference.  Also, when you loop as much as
calculating some of these indicators requires, any extra commands
equal noticeable performance hits.  

I'm convinced that when performance becomes a priority, this issue
will improve.  There are just too many other things TS is doing wrong
now for them to have top coders working on this now.  After all, what
good is drawing a chart faster if it looks wrong? <g>

       8