On 2 Jun 2003 at 11:42, Richard H. McCullough wrote:

> 
> I modified MKE so that it does not generate 
> the very large sets (72,000 & 3,300 members). 
> The results (lines/15 min.) are shown below.
> 
> I conclude that
> 1. WinXP task manager is a serious problem
> 2. very large sets are a serious problem
> 3. some other factors (as yet unidentified)
> cause significantslowdowns.

My machine is like yours - 2Ghz, 512M, WinXP.
I have run tables over 2 million items with no slow-down - 
as long as the data set is ENTIRELY in RAM 
AND all string keys >10 chars differ in the first 10 chars.
(latest version of Unicon - older ones may be slower @ 2M)
Adding 50000 string keys that all have the same first 10 chars 
to a table is 200 times slower than normal (no swapping)
Adding 60000 almost always made the test exit. (that's a bug)
(one successful run was 450 times slower)
(a recent CVS MinGW build quit before 45000)
If CPU% hits 90-100% with no results,
this could be the problem.
If it hits the swapfile, it could easily be 2000 times slower.

In WinXP task manager, if the process list does not have 
the working set & VM size & page fault columns, turn them on.
Sort by working set. If you know what services are safe to stop
that you don't need, stop them to save RAM.

The low CPU % is almost always due to thrashing - not 
enough RAM. Use MMC to monitor % disk time for the 
physical drive with your swap file.
If it hits 100%, the CPU is most likely waiting for it.
(if you have 2 drives, make 2 swap files.)
If your program is not doing I/O,
A 1% CPU time & 100% disk time indicates ~100X slow-down
due to swapping.
2% CPU = 50X slow-down, 20% CPU = 5X slow-down
(for 1 CPU, no hyperthreading, no background disk or CPU load) 

Dr. Jeffery is right - virtual memory 
and traditional garbage collecting do not get along - 
100 x slow-down is possible.
I did a Unicon runtime that logs garbage collection time;
the 90% threshold to up-allocate was the main culprit,
as Dr. Jeffery said.



-------------------------------------------------------
This SF.net email is sponsored by:  Etnus, makers of TotalView, The best
thread debugger on the planet. Designed with thread debugging features
you've never dreamed of, try TotalView 6 free at www.etnus.com.
_______________________________________________
Unicon-group mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/unicon-group

Reply via email to