Re: Mersenne: Intel IA64 AMD Sledgehammer
On 30 Oct 2001, at 10:37, Lars Lindley wrote: Are there plans for making mprime and prime95 capable of using 64bit processors like IA64 and AMD Sledgehammer? I know that these processors are able to run 32 bit code but will there be a 64bit optimized version of mprime/prime95? IA64 may run IA32 code but I seem to remember this is far from optimal. Hammer (which AFAIK exists only as vapourware) is supposed to run IA32 code efficiently. What would we expect to gain from extended addressing, or 64 bit integer operations? Probably not much for LL testing ... that's using double-precision floating-point, which is already native 64+ bit hardware in IA32. Trial factoring would almost certainly speed up a fair bit with 64 bit integer hardware operations. There may be mileage in further optimization for Itanium (IA64) so as to take better advantage of the parallel execution paths than the current scalar code. However Itanium processors are still a frightening price; I can't see them becoming consumer items in the near future. Also, Glucas already works very well on Itanium systems! Regards Brian Beesley _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Re: SV: Mersenne: number of processors participating
On 29 Oct 2001, at 19:37, John R Pierce wrote: I've had similar problems with a few other multimedia sorts of junkware. Near as I can tell, some of these things put their video or animation thread at Idle_Priority+1 or something, and it gets eaten alive by Prime95. Isn't it the old problem - no matter what priority a process is running at, unless it's interrupt driven it won't preempt a process running at a lower priority. The problem here is that the multimedia stuff wants to do a very little work but very often. It gets slowed down because Prime95 hangs on to the processor until its timeslice expires - it almost never has to wait for some external event. Ideally the multimedia stuff would be driven by timer interrupt. But for some reason (maybe something to do with there being a limited number of timer channels, and those having rather poor resolution) this approach seems to be quite rare on PC systems. One way to improve the performance in these circumstances is to reduce the minimum timeslice for low-priority processes. This will cause the task scheduler to be busier and therefore reduce the overall performance to some extent, but multimedia type applications will coexist much more happily with compute-intensive tasks if this is done. Sorry, I have no idea how to do this, or even whether it is possible, in any of the versions of Windows. The linux 2.4 kernel does this almost automatically, by having a much smaller minimum timeslice for idle-priority processes than for processes running above idle priority. (The timeslice is reduced again for processes running at unusually high priority, so that they can't hog the whole system quite so easily.) I believe the timeslice parameters are tunable (without having to recompile the kernel), but I have no personal experience of actually doing this. Another other way to fix the problem is to have the compute- intensive process voluntarily relinquish its timeslice at intervals which are much shorter than the minimum timeslice (which is typically of the order of 200 ms). This reduces the efficiency of the compute-intensive task to some extent but does make it coexist better. I suppose it would be possible to build this into Prime95; if this is done I would like options to be multimedia friendly or optimally efficient - probably the best way to implement would be to have the code contain the relevant system calls but to NOOP over them if efficiency is demanded. The remaining problem with this approach is that how often you would want to make these system calls would depend very heavily on the processor speed. Relinquishing the timeslice very frequently would enable even slow systems to run multimedia pretty seamlessly, but at a heavy cost on all systems. Placing the system calls in a position where they would be effective but not too costly, across a large range of processor speeds and a large range of FFT run lengths, would not be a trivial task. Regards Brian Beesley _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Mersenne: Re: number of processors participating
On Tue, Oct 30, 2001 at 10:00:07PM -, [EMAIL PROTECTED] wrote: which are much shorter than the minimum timeslice (which is typically of the order of 200 ms). 200ms? Wouldn't this be an error? I can't really imagine that one would typically have only five time slices per second :-) /* Steinar */ -- Homepage: http://www.sesse.net/ _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Re: SV: Mersenne: number of processors participating
I've had similar problems with a few other multimedia sorts of junkware. Near as I can tell, some of these things put their video or animation thread at Idle_Priority+1 or something, and it gets eaten alive by Prime95. Isn't it the old problem - no matter what priority a process is running at, unless it's interrupt driven it won't preempt a process running at a lower priority. ... process and thread dispatching in MS Windows IS interrupt driven. Anything that causes a thread or process thats waiting to become ready will cause it to immediately dispatch if its the highest priority ready process, the system doesn't wait for the next major quantum tick. Multimedia stuff is either waiting on sound buffer events, or multimedia timer events (which have 1mS resolution) or disk IO buffer events, or software semaphore events, all of which are interrupt driven and will cause an immediate dispatch. near as I can guess, the issue here is that Prime95 is running a few priority notches above idle and when another process tries to run at a lower priority it will stall behind prime95. _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
RE: SV: Mersenne: number of processors participating
One way to improve the performance in these circumstances is to reduce the minimum timeslice for low-priority processes. This will cause the task scheduler to be busier and therefore reduce the overall performance to some extent, but multimedia type applications will coexist much more happily with compute-intensive tasks if this is done. Sorry, I have no idea how to do this, or even whether it is possible, in any of the versions of Windows. There is a program that can set the quanta for programs... let me find that durned thing... Aha. http://www.sysinternals.com/ntw2k/freeware/frob.shtml Good old sysinternals... they have the neatest tools. Apparently that's just for NT4 machines (I think...). For Win2K (and presumably XP?), they have another page that tells you about the settings on there, and to wait for a new version of Frob that works with win2k. http://www.sysinternals.com/ntw2k/info/nt5.shtml Aaron _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
SV: SV: Mersenne: number of processors participating
Ups, by help from Brian Beesley and a little work with the time= I have it working now. I think it was my old paranoia from a time when I was not running the servers alone - I wouldn't let anyone know that a program like prime95 was active. Now I don't care as I have nobody but users to face. Thanks to all. Happy hunting tsc -Oprindelig meddelelse- Fra: George Woltman Sendt: ma 29-10-2001 22:47 Til: Torben Schlüntz Cc: Emne: Re: SV: Mersenne: number of processors participating Hi, At 10:01 PM 10/29/2001 +0100, you wrote: I would like to use the servers; but I haven't been able to persuade George to make a Quit function like quit_at: 06:00 to terminate the program when users arrives and optimum performance is needed Look in readme.txt for the Time= entry in prime.ini This feature can be used to make prime95 go dormant at a specified time. Hope that helps, George _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
RE: SV: Mersenne: number of processors participating
Aaron Blosser wrote: Good old sysinternals... they have the neatest tools. Damn straight! I've been using (and loving) PageDefrag since I stumbled on that site. A few other gems have since made their way onto my system... Rick. -+--- [EMAIL PROTECTED] http://www.alienshore.com/seeking/ _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Re: Mersenne: Re: number of processors participating
As I recall, 200 was the default quantum on a vax 780 20 years ago. Even 10 years ago, interactive response could be sped up a lot by cranking quantum down to single digits. At 11:28 PM 10/30/2001 +0100, Steinar H. Gunderson wrote: On Tue, Oct 30, 2001 at 10:00:07PM -, [EMAIL PROTECTED] wrote: which are much shorter than the minimum timeslice (which is typically of the order of 200 ms). 200ms? Wouldn't this be an error? I can't really imagine that one would typically have only five time slices per second :-) /* Steinar */ -- Homepage: http://www.sesse.net/ _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Thrashing (was Re: SV: Mersenne: number of processors participating)
What will really slow a workstation or server down is running short of RAM. These days the working sets are getting appreciable as the exponents increase. NT scheduling will wake up the service version of ntprime every second I think and give it at least one quantum. If some more essential service or application needs nearly all available RAM for its working set, and the working set of ntprime is big enough it gets paged out, the disk thrashes wildly and performance can suffer greatly for both the ntprime service and the other service or application, even while the ntprime service only gets a percent or two of cpu time. This is not just a characteristic of NT, but a general property of virtual memory operating systems; eventually it's just too little ram or too much demand, leading to performance decline. Ken At 05:05 PM 10/29/2001 -0800, Aaron Blosser wrote: Still the only time I've ever seen Prime95/NTPrime slow down a system is when I was doing some Netmeeting video conferences. _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers