QUESTION
> we intent to buy a AlphaStation or AlphaServer. But this morning i got
> the Summary about "AlphaStation 4/233 vs. 3400" Performance describing
> the outperforming of the newer 4/233-machine by the older 3400.
> There are a lot of factors which determinates the performance of a
> special application on a special platform. Dave Cherkus answer was
> that the main memory bandwith may cause the difference.
> My question: Where can i find informations on the bandwith of main memory
> for Alpha 3000 AXP and the newer Alpha Stations/AlphaServers.
> Is there a file/directory on gatekeeper to search for?
THANKS
Dave Cherkus cherkus_at_UniMaster.COM
Bill Bathurst bathurst_at_corona.pmel.noaa.gov
Jyotinath jyoti_at_qcav01.enet.dec.com
Special thanks to Mr. Neidecker-Lutz from CEC Karlsruhe (Digital, Germany),
who make it possible to run tests on several 21064-machines.
ANSWERS
...
Yes, my answer was that in the application this person described, I
thought main memory bandwith may cause the difference. In some other
applications, the CPU clock rate might be more important.
...
Try
http://www.digital.com/info/misc/axp-ws-summary.abs.html
which has info about the older systems. I don't have a source
for the newer ones.
Also, I recommend the 'stream' benchmark, available via ftp from
perelandra.cms.udel.edu, or the lmbench1.0 benchmark posted to
comp.sources.unix, etc.
Dave Cherkus ----- UniMaster, Inc. ----- Contract Software Development
---------------------------------------------------------------------------
...
I have found NFS 3.0 is a big improvement over machines that don't have it.
Right now only DEC and Sun have NFS 3.0. For us it has been a 4 fold
increase in NFS performance over NFS 2.0 machines.
Bill
---------------------------------------------------------------------------
...
Anonymous ftp to gatekeeper.dec.com; look in
pub/DEC/DECinfo/performance/sys directory for
a range of server and workstation (OSF/1, VMS
and NT) performance flashes and briefs (less
detailed and more detailed).
/Jyoti
---------------------------------------------------------------------------
The www-page axp-ws-summary.abs.html (see Dave) is a good summary for the Alpha
3000 AXP Family. (gatekeeper:/pub/Digital/info/misc/axp-ws-summary.*) For
special machines may be of interest (technical summaries):
Alpha 3400 AXP ../technical-overview/dec-3000-400.*
Alpha 3500 AXP ../technical-overview/dec-3000-500.*
AlphaServer 2000/2100 ../technical-overview/EC-N3822-10.*
Memory Specifications (see axp-ws-summary):
300 300L 400 600 500 500X 800
Bus Width [bits] 64 64 256 256 256 256 256
Bus Speed [MB/s] 66 66 106 114 114 114 114
The 2000/2100s has a 256-bit interface to DRAMs, the AlphaStations/Servers
are 64 bit width, i have been told.
But the Memorybuswidth is only one, may be an interesting, factor for
calculating the performance of the applications you want to run. Think
about the cache, the cycles, i/o-, cpu-, network-intensive appl. (see
Bill), the different benchmarks Digital presents us (see Jyoti), and so forth.
The best benchmark is to run your own application on the platform you are
interested in. So I asked for a test and Mr. Neidecker-Lutz was able to do it.
The aim wasn't to do extensive benchmarking, but to get a feeling about the
different behaviour of the machines running the same programme.
We configured our program (aiolos, a numerical simulation fortran program) to
need ~40 MB memory; "size aiolos.exe" gives:
text data bss dec hex
344064 40960 38019968 38404992 24a0380
(OSF/1 V1.3A, DEC Fortran 3.3)
Tests done from Mr. Neiderdecker-Lutz
Machine Mem OSF/1 R. Output from time-command (csh, without elapsed time)
3300X 64 V3.0 347 14496.365u 20.295s 99.3% 0+352k 241+1159io 341pf+0w
3500 96 T3.2 185 12018.024u 111.027s 95.7% 0+375k 168+1026io 68pf+0w
3600 128 T3.0 358. 10357.596u 65.471s 99.8% 0+376k 82+879io 423pf+0w
200 4/166 160 V3.0 358 12731.368u 10.277s 99.3% 0+376k 0+1040io 336pf+0w
200 4/233 96 T3.2 185 9069.328u 7.418s 99.6% 0+375k 2+1io 0pf+0w
Fortran RTL Vers. DFARTL356 on all machines
Own Tests
Machine Mem OSF/1 R. Output from time-command (csh, without elapsed time)
3300X 160 V3.0 347 1)
3500 192 V1.3A 185 11344.97u 65.471s 99.8% 0+376k 82+879io 423pf+0w
Comparison (aiolos based on user-cpu-time)
Maschine Cache SPECint92 SPECfp92 AIOLOS
-----------------------------------------------------------------------------
3000/300X 0.25 1.00 1.00 1.00 D
O 1)
3000/500 0.5 1.00 1.27 1.20 D
1.28 O
3000/600 2.0 1.35 1.63 1.40 D
200 4/166 0.5 1.38 1.34 1.14 D
200 4/233 0.5 1.87 1.83 1.60 D
D Digital, O Own,
1) we will do a testrun as soon as possible (3-5 days) and place the results
in ftp.ivd.uni-stuttgart.de:/pub/aom_sum/3300X.result (129.69.23.6)
I wanna say it's a good thing to get the performance charts from Digital.
Adding the offering of testruns is a policy to gain confidence of a customer.
Claus
#############################################################################
#address: University of Stuttgart #
# Institute of Process Engineering and Power Plant Technology (IVD) #
# Claus Nagel, Pfaffenwaldring 23, D-70550 Stuttgart . Germany #
#phone: 49 711 685 -3751, -3563, -3487 _ - |\ #
#fax: 49 711 685 3491 - | \ #
#email: nagel_at_ivd.uni-stuttgart.de | \ */ #
# claus.nagel_at_rus.uni-stuttgart.de .__|___\|_. #
########################################################~~~~\_______/~~~#####
Received on Sun Jan 15 1995 - 13:00:58 NZDT