Author: Vincent Diepeveen
Date: 12:22:19 08/03/03
Go up one level in this thread
On August 03, 2003 at 14:49:12, Eugene Nalimov wrote:
>On August 03, 2003 at 11:12:15, Vincent Diepeveen wrote:
>
>>On August 03, 2003 at 06:14:06, Bo Persson wrote:
>>
>>>On August 02, 2003 at 18:22:58, Vincent Diepeveen wrote:
>>>
>>>>On August 02, 2003 at 09:15:26, Bo Persson wrote:
>>>>
>>>>>On August 01, 2003 at 14:13:18, Kim Roper Jensen wrote:
>>>>>
>>>>>>On August 01, 2003 at 14:00:18, Eugene Nalimov wrote:
>>>>>>
>>>>>>>I see quite different result on AMD64 when Crafty is compiled by 64-bit Visual
>>>>>>>C. :-)
>>>>>>>
>>>>>>>Thanks,
>>>>>>>Eugene
>>>>>>>
>>>>>>
>>>>>>Sorry for asking but what did you see ?? :)
>>>>>
>>>>>
>>>>>A non-disclosure agreement in his employment contract? :-)
>>>>>
>>>>>I'm sure we will see the numbers about 2 seconds after the compiler is released.
>>>>
>>>>Not really. What you need first is a windows version that can run the stuff 64
>>>>bits native at opteron.
>>>
>>>Yes, of course. Do you care to make a guess on what MS is using their Opteron
>>>compiler for. :-)
>>
>>I guess they are working hard and basic problem is not only that m$ is 32 bits
>>in some respects (file systems and such already long period ago 64 bits in 1995)
>>but especially that their kernel stuff is written in assembly.
>
>This is definitely so, especially if you take into account that NT/2k/XP
>variants were commercially sold not only for x86, but for PowerPC, MIPS, Alpha,
>and IA-64 CPUs. Of course MS wrote 5 kernels in the different assembler
>languages...
It is hard for me to understand why so much is in assembler and why there
seemingly is no x bits 'C' version which they can compile on any hardware with
little effort.
All you need then is a good compiler and perhaps just a bit of handtuning here
and there.
Even a way slower C version of such an OS would kick of course incredible butt
already for the different testsets. Now and in the future.
In the meantime what i'm using is:
bash-2.05$ uname -a
Linux a1.teras.sara.nl 2.4.20-sgi220r3 #1 SMP Wed Jun 4 16:43:37 PDT 2003 ia64
unknown
bash-2.05$ cat /proc/sys/kernel/shmmax
90387103744
This REMINDS me of a problem i have with windows. In order to get for different
processes pointers in shared memory at the same virtual adress space i use next:
TreeFileMap = OpenFileMapping(FILE_MAP_ALL_ACCESS,FALSE,
"DiepMapFile_Tree");
if( TreeFileMap == NULL )
mprint("Fatal Error Mapping File: 'tree' for proces %i\n",
ProcessNumber);
tree = (TreeInfo *)MapViewOfFileEx(TreeFileMap,
FILE_MAP_ALL_ACCESS,0,0,0,(LPVOID)0x01000000);
and of course createfilemapping to create it. The problem is that last parameter
in mapviewoffileex. I put it hardcoded at a certain adress in the memory. What i
fear is that some software like the JRE and some databases (oracle?) are using
similar adresses to map stuff into the same adress space.
If they take my adress then i have bad luck simply.
In Linux/IRIX/UNIX there is a way around that because i use ftok to get an
unique Key not yet in use by other software. Then shmget/shmat puts
automatically the allocated memory in the same adress space for the processes.
Any way to prevent other processes taking care diep can't start parallel?
>Thanks,
>Eugene
>
>>That's one of the reasons why the kernel is much faster practically than linux
>>for applications (if i put of old diep versions which aren't NUMA 2 processes to
>>search at a single cpu machine then it runs *way* faster on NT, about factor 2
>>to 4 than under linux kernel).
>>
>>So using their own compiler for their own kernel is not exactly what they can
>>do.
>>
>>A big challenge of opteron is that it is NUMA.
>>
>>I am sure m$ will be very happy supporting the AMD platforms. Competition in the
>>hardware branche is good for microsoft. So in the highest levels of the
>>organisations the x86-64 platforms will have a lot of support.
>>
>>Even if microsoft wouldn't want to support it, they still MUST support it
>>because x86 is history within a few years and we all will be running x86-64
>>platforms only. Either with intel OR amd sticker.
>>
>>I guess even the biggest cpu amateurs will understand by now that intel is
>>developing their own x86-64 processor generations. Perhaps even giving it a 'p4'
>>sticker though it's an entirely new core.
>>
>>The only thing we do not know is *when* they will release their x86-64 cpu's.
>>AMD simply has advantage there now.
>>
>>We can very shortly describe the x86-64 architecture. Cheap. High clockable and
>>superior to everything out there including itanium. Especially superior to
>>itanium.
>>
>>I'm running at a cpu or 64 now (itanium2-madison 1.3Ghz 3MB) and they are great
>>for the highend but a joke even when compared to x86 for the average user.
>>
>>Intel plans to mass produce itaniums for the 'low-end' market have been put into
>>the fridge a long while ago. The only reason intel is continuing this processor
>>now (seemingly) is because they probably can't go back. Or perhaps they wait
>>until they have x86-64 cpu's available.
>>
>>How can GCC 2.96 without profile recompilation be just 15% slower than intel c++
>>7.0 using profile information (prof_use) at the itanium platform?
>>
>>The problem of the itanium platform is they can't clock it high despite working
>>for years already at that problem, it is too expensive, and it is impossible to
>>write software for it.
>>
>>Even the current generations of supercomputers with itaniums that get delivered
>>are missing major software support for it. Like crucial fortran libraries.
>>
>>This where > 60% of the total system time of supercomputers goes to gflops used
>>by fortran libraries.
>>
>>If you add up the picture then it is a matter of time before the x86-64 will
>>dominate everything.
>>
>>However it is sad to realize that most likely the linux world will be too late
>>again. Despite that microsoft must convert their assembly libraries to new
>>opteron assembly, we know they must be very far already with that conversion.
>>
>>This where the linux plans to write a NUMA kernel for kernel 2.6 have been just
>>defined a few weeks ago. Not to mention the years it will take to carry out an
>>effective implementation.
>>
>>For those who still do not understand what i'm talking about. When you have more
>>than 1 opteron processor, so a dual opteron or even more processors, then it is
>>of crucial importance that the kernel can run locally on each kernel for the
>>jobs that can be done locally.
>>
>>Local memory at opteron is way faster than global memory.
>>
>>In fact the opteron asks for a cc-NUMA operating system which until recently was
>>only getting used by very big non-real time supercomputer systems.
>>
>>I would not be surprised if microsoft is years sooner in releasing a version
>>that works well than the linux community, despite that everyone will understand
>>that bugfixing assembly code is a lot harder than fixing a bit of C code.
>>
>>Best regards,
>>Vincent
>>
>>>
>>>Bo Persson
>>>bop2@telia.com
This page took 0.01 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.