[Wien] performance problem using ifort + Goto's lib

[Sun Yiyang]

Since the hardware is almost the same (P4C with 800FSB and 512M L2 Cache, 
DDR400 Dual-ch RAM, Intel 865 Chipset), I would expect the performance difference
is on the software side.
 
One point mentioned by Jose may make sense, that is the hyper-threading. 
Intel knows this technique best. That may explain why MKL yields better performance
than Goto's lib. Unfortunately, I failed to use MKL6.1 together with all versions of ifc.
Sometimes it's due to compilation error, sometimes run-time error as I said in my original
thread. 
 
Anyway, thanks for all the replies. I'll give up digging out the last bit of performance and 
start to work with WIEN2k!
 
Sincerely,
Yiyang
 
-----Original Message----- 
From: wien-admin at zeus.theochem.tuwien.ac.at on behalf of Jose Mestnik Filho 
Sent: Tue 3/30/2004 9:34 PM 
To: wien at zeus.theochem.tuwien.ac.at 
Cc: 
Subject: Re: [Wien] performance problem using ifort + Goto's lib



	Dear Wien users,
	This mater seems to be of great interest to the community. So let me be
	more precise:
	
	Mother board: Intel D865 Perl, 800MHz, dual channel memory
	memory: 1GB, 400MHz, DDR, two memory boards 512Mb each. This is
	necessary to take the advantage of the dual channel. Expandable up to 4Gb.
	Processor: Pentium 4HT, 3.2 GHz, 512Kb cache
	
	The benchmarks for the test_case, for various compilation options were:
	ifc6.0 + mkl6.1: 226s (I did not succeed to compile with ifort8.0)
	ifc6.0 + ATLAS(1 thread) : 247 s
	ifc6.0 + ATLAS (2threads): 362 s (I would expect it to be better,
	because of HT, but, as seen, it is not)
	ifc6.0 + goto p4_512-pr0.94: 317s
	ifc6.0 + goto p4_512-r0.94: not working
	
	Interesting is that these tests were made with NMATMAX=5000. In the case
	NMATMAX=10000, the timings increase to 234s with ifc6.0 + mkl6.1
	
	Just the "crude" operating system is running. No X11 or whatever else.
	
	Regards,
	Jose Mestnik Filho
	
	Peter Blaha wrote:
	
	> P4 (3.2 GHz) is probably not enough to specify a system. Motherboard, memory,
	> and specific version of the processor (cache, hypertrading,...) may change
	> the performance by these 10-15 %.
	>
	> The "228" seconds were reported by
	> Lutz Rauscher" <rauscher at qms.cache.waseda.ac.jp>
	> and confirmed by
	> Jose Mestnik Filho <jmestnik at baitaca.ipen.br>:
	> "The computers are pentium 4HT, 3.2 Gh, dual
	> channel memory, 1 Gb memory each, expandable until 4Gb. The test_case
	> matches exactly with the known benchmarks (226 s)."
	>
	>
	>
	>>>P4, 3.2 GHz, dual channel mem.   258 sec    ifc7, mkl6
	>>>P4, 3.2 GHz, 400MHz dual ch.mem. 228 sec    ifort8.0, mkl6.1
	>>>
	>>>Not really a clean comparison (ifc vs ifort as well), but this gain of
	>>>30 s is less than the estimated gain of goto over mkl 6.0. If this
	>>>conclusion is true (do other users find the same?), then this explains
	>>>your above question.
	>>>
	>>
	>>I did the original benchmark on our workstation that yielded the 258 sec
	>>result.  This benchmark actually involved the ifc 7.1 compiler and the
	>>mkl 6.1 libraries.  I have since applied the latest update to the mkl
	>>libraries and recompiled with the ifort 8.0 compiler and I find no change
	>>in the benchmark time - with some optimization, I now get a benchmark
	>>time of 256 sec, and if you kill absolutely ever process, then you can
	>>trim it down to 254 sec.  Compiling with the Goto blas routines also
	>>didn't seem to make much of a difference (on the order of seconds).
	>>I think this result is consistent with benchmark time of 258 secs that
	>>was quoted earlier in this thread.
	>
	>
	>                                       P.Blaha
	> --------------------------------------------------------------------------
	> Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
	> Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
	> Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
	> --------------------------------------------------------------------------
	>
	> _______________________________________________
	> Wien mailing list
	> Wien at zeus.theochem.tuwien.ac.at
	> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
	>
	>
	
	--
	----------------------------------------------
	Jose Mestnik-Filho <jmestnik at net.ipen.br>
	Instituto de Pesquisas Energeticas e Nucleares
	Research Reactor Center - CRPq
	Po.Box. 11049
	05422-970 Sao Paulo SP Brasil
	----------------------------------------------
	
	
	_______________________________________________
	Wien mailing list
	Wien at zeus.theochem.tuwien.ac.at
	http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
	

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/ms-tnef
Size: 8838 bytes
Desc: not available
Url : http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20040331/000c3baf/attachment.bin