Benchmark of Older IBM Mainframes vs. Intel Based
Ivan Flores
I need to compare in the same (or approximately) the same terms
older IBM computers like IBM 4381 or 3090 against newer Intel based
processors 486 and Pentiums. I have found lots of info regarding
the Intel processors and different units but nothing about the IBMs.
Additionally. Does any body know the MIPS mark of the VAX 11/780?
Any unit of measure will be helpful: SPEC, FP, NFS, or MIPS.
Thanks in advance.
Ivan Flores (flo...@cgs.edu)
Alex Mericas
>I am looking for benchamarks in the net about older systems.
>Any unit of measure will be helpful: SPEC, FP, NFS, or MIPS.
It is very important to remember that most INTEL systems are single user systems while
most 370/390 systems are multi-user. INTEL based operating systems are optimized to
providing the maximum amount of system resourses to the end user, while 370/390
operating systems are optimized to maximize system throughput. So it may be possible
for a Pentium to execute a loop more times than a single CICS or CMS user, the
CICS or CMS user is sharing the system with 1000's of other programs.
The TPC benchmarks look at throughput on a system level. TPC-A is not very
representative of real workloads, but can provide some interesting insites into
system capacity. I don't have current data, but from results as of 12/15/93:
IBM 9121-511 Running TPF - 3504 TPS
IBM 9121-742 Runnins IMS - 1427 TPS
COMPAQ 5/66 Running SCO + Oracle7 - 242.23 TPS
Alex Mericas Westlake Performance mer...@vnet.ibm.com
Jeffrey Reilly
>In <3nfii3$s...@jaws.cs.hmc.edu>, Ivan Flores <flo...@cgs.edu> writes:
>>I am looking for benchamarks in the net about older systems.
>>....
>
>representative of real workloads, but can provide some interesting insites into
>system capacity. I don't have current data, but from results as of 12/15/93:
>
> IBM 9121-511 Running TPF - 3504 TPS
> IBM 9121-742 Runnins IMS - 1427 TPS
> COMPAQ 5/66 Running SCO + Oracle7 - 242.23 TPS
>
Do those machines meet the "older system" requirement asked for by the
original poster? The Compaq system appears older than the IBM systems (if
one assumes that the TPC results were provided at product announcement; I
do not know enough about IBM systems to know.)
And, of course, there is the second metric that TPC requires users of their
metrics to provide:
System tps-A $/tps-A Published Result status
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
ES/9000 Model 511 TPF 3504.9 $8,310 12/93 Current as of 2/95
ES/9000 Model 742 IMS 1427.1 $13,586 12/93 Current as of 2/95
Compaq SysPro/XL 242.2 $5,130 8/93 Withdrawn by vendor
12/94
Jeff
Associate Editor, SPEC Newsletter
Jeff Reilly | "There is something fascinating about
Intel Corporation | science. One gets such wholesale returns
jwre...@mipos2.intel.com | of conjecture out of such a trifling
(408) 765 - 5909 | investment of fact" - M. Twain
Nick Maclaren
|> I am looking for benchamarks in the net about older systems.
|> older IBM computers like IBM 4381 or 3090 against newer Intel based
|> processors 486 and Pentiums. I have found lots of info regarding
|> the Intel processors and different units but nothing about the IBMs.
This seems a little bizarre. For a short time, we have a 3084, and
I have run some benchmarks on it. However, it is NOT commensurate
with Intel chip systems. The two types of system are so different
that I can imagine no useful purpose in benchmarking them against
each other. I don't regard the production of marketing hype as a
useful activity :-)
Timing a particular application on the two types of system may well
be useful, but that isn't what you asked about.
|> Additionally. Does any body know the MIPS mark of the VAX 11/780?
|> Any unit of measure will be helpful: SPEC, FP, NFS, or MIPS.
Very approximately 1 MIPS.
Nick Maclaren,
University of Cambridge Computer Laboratory,
New Museums Site, Pembroke Street, Cambridge CB2 3QG, England.
Email: nm...@cam.ac.uk
Tel.: +44 1223 334761 Fax: +44 1223 334679
Ivan Flores
>
> In article <3nfii3$s...@jaws.cs.hmc.edu>, Ivan Flores <flo...@cgs.edu> writes:
> |> older IBM computers like IBM 4381 or 3090 against newer Intel based
> |> processors 486 and Pentiums. I have found lots of info regarding
> |> the Intel processors and different units but nothing about the IBMs.
>
> This seems a little bizarre. For a short time, we have a 3084, and
> I have run some benchmarks on it. However, it is NOT commensurate
> with Intel chip systems. The two types of system are so different
> that I can imagine no useful purpose in benchmarking them against
> each other. I don't regard the production of marketing hype as a
> useful activity :-)
We are studying a case in which an IBM 3090 was downsized to a
client/server platform, where the server was a 486-25Mhz and SCSI
disk drives with nearly 5GB, plus LAN adapters. This happened in
1990 in the Pacific Misile Test Center in Point Mugu, California.
The application running in this system was one of the 3 DBs under
the Airborne Weapons Information System.
My point here is measuring only processing speed. Speed of channels,
disk drives, or other external devices will be analyzed separatelly.
At the end we will evaluate the integration feasibility. We are also
considering the type of applications. Commercial performance, we
understand has very complex components and has to be measured by
using TPC-C type of bechmarks.
Ivan Flores
Hugh LaMaster
jwre...@mipos2.intel.com (Jeffrey Reilly) writes:
|> >In <3nfii3$s...@jaws.cs.hmc.edu>, Ivan Flores <flo...@cgs.edu> writes:
|> >>I am looking for benchamarks in the net about older systems.
Without knowing exactly which 3090 we are talking about
here, it is pretty difficult to even extrapolate vaguely.
For example, looking at the -711 based ES/9000 models,
from an ES/9000 -330 to a -9X2 represents about
a 25:1 or 50:1 capacity improvement. The range
for the -511 based processors is smaller, but
still substantial.
|> > IBM 9121-511 Running TPF - 3504 TPS
|> > IBM 9121-742 Runnins IMS - 1427 TPS
|> > COMPAQ 5/66 Running SCO + Oracle7 - 242.23 TPS
:
|> System tps-A $/tps-A Published Result status
|> - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
|> ES/9000 Model 511 TPF 3504.9 $8,310 12/93 Current as of 2/95
|> ES/9000 Model 742 IMS 1427.1 $13,586 12/93 Current as of 2/95
|> Compaq SysPro/XL 242.2 $5,130 8/93 Withdrawn by vendor
Since a 511 is about 1/4 of a 742, the choice of software
obviously makes a difference. It appears that IMS is
about 10X as consumptive of resources as TPF. The much
more lightweight [apparently] TPF is probably not the I/O hog
that IMS (or Oracle) would be. I doubt if it has the same
capabilities, either (e.g. to do a TPC-B and/or TPC-C).
In another thread, pe...@wimsey.com (Patricia Eddy) writees;
|> I want to be able to select a RISC box that will give me the mainframe
|> processing times. I have a SAS procedure ( sort and merge) that we use as
|> a benchmark. This consistently takes 5 minutes of CPU time on the mainframe.
|> I want a UNIX box that will take NO MORE THAN 5 minutes of CPU time to so the
|> same procedure. I know this is crude but I'm trying to get *minimum*
|> performance ratings based on what is acceptable processing time. This has
|> proved almost impossible for me to find out on a mainframe ( even from the
|> manufacturer).
Which of the RISC systems does SAS run on? Why not
just run your benchmark on one of their systems?
Sales reps. can usually arrange this.
|> Here is what I have:
|>
|>
|> Amdahl 5995-4570M 214 MIPS (GArtner) 512 MB main memory, 512 MB expanded mem.
|> Amdahl 5995-1400A 108 MIPS(Gartner) 448 MB main memory, 320 MB expanded mem.
It is interesting to think about what these "MIPS" ratings
mean. Below, it appears that a comparable RISC-based
system might need 1200-1600 SPECints (just taking
each processor * number of processors)? Does this
mean that 1 "Gartner group" MIPS == 5.5-7.5 SPECint "MIPS"?
On single CPU-bound programs, I would guess that it is
more like a 2:1 ratio. [Does anybody have any numbers?]
My guess is that the RISC servers are burning a lot
more cycles doing I/O. Why don't they start putting
fancy channels on RISC-based servers?
|> This is the only rating I have:
|>
|> Amdahl 5995-4570M TPC-B 1116.4 tpsB
What software was used for this rating? IMS? Oracle?
How many processors is the 5995-4570M? (I think it is
four?)
If you need around 1100 tpsB, there are only a few systems
with that kind of performance. [I'm not clear from your
posting if you are trying to replace the Amdahl, or merely
pull one SAS application off ...] 1000+ tpsB is probably
a good definition of "mainframe" performance, though.
Just to move that one SAS application off the Amdahl
could require nothing more than a modest RISC-based
server system.
DEC has a number of Alpha models (I'm not sure of the
new model numbers) using 250, 275, and 300 MHz Alphas
with pretty fast single-CPU SPECint92 performance ratings.
For example, a DEC Alphaserver 2100 4/275 can be had in a
decent 1-4 processor configuration for $50-$200K list.
[It all depends on what you need in the configuration.]
For mid-level performance, these systems seem like
pretty good deals. Does SAS run on OSF/1 on the Alpha?
If you need the throughput implied by the 1000+
TPC-B numbers, it is going to cost you a bit more.
Amdahl, DEC, Tandem, SGI, Sequent, and Cray have all published
tpsB numbers that large. The highest that I know of is a
Cray SPARC-based server (the CS6400). It is a 32-processor
(can go to 64 processors) SPARC-based system which costs
in the $2-$3 Million range. Amdahl has a similar SPARC-based
system. I don't have exact numbers, but you are probably
looking at something like a 16-20 processor version of the
SGI Challenge L, Cray CS6400, Amdahl SPARCsummit 6400E,
Sun SPARCcenter 2000E, or perhaps only a 4-8 processor
DEC 8200/8400. DEC has some new systems based on the
300 MHz Alpha, the new DEC AlphaServer 8200/8400 lines.
They seem to have comparable performance, but with the
advantage of having fewer, faster, CPUs. AFAIK, they
haven't published TPC-B and/or TPC-C numbers yet.
[I wonder if DEC OSF/1 can be booted *bigendian*?]
I think all these systems, when configured with the
necessary processors, memory, disks, controllers, etc.,
would all be in the $1-$3 Million range, and I'm sure
your salesrep would be happy to talk to you about it.
Anyway, all this is by way of saying that, yes, it
appears that the latest generation of RISC-based
*high-end* servers are pretty comparable to "mainframes"
in performance. A few companies, like Cray and Amdahl,
are beginning to make claims about mainframe-like RAS
via improvements in hardware and/or software as well.
--
Hugh LaMaster, M/S 233-9, UUCP: ames!lamaster
NASA Ames Research Center Internet: lama...@ames.arc.nasa.gov
Moffett Field, CA 94035-1000 Or: lama...@george.arc.nasa.gov
Phone: 415/604-1056 #include <std_disclaimer.h>
Hugh LaMaster
>> Why don't they start putting fancy channels on
>> RISC-based servers?
BTW, I am aware that one answer to this question is that when
you consider the cost of bringing a system to market, it is
cheaper just to put 2-4X CPUs on a large system. This is OK,
but fancy channels can save some memory B/W too, which is the
bottleneck on most full-up server systems. Different approaches.
The proof of the pudding in this case is that, apparently,
high-end RISC-based Unix servers *are* getting results
comparable to the mainframes now. Now, about R-A-S...
Nick Maclaren
Hugh LaMaster <lama...@viking.arc.nasa.gov> wrote:
>lama...@viking.arc.nasa.gov (Hugh LaMaster) wrote:
>
>>> Why don't they start putting fancy channels on
>>> RISC-based servers?
>
>you consider the cost of bringing a system to market, it is
>cheaper just to put 2-4X CPUs on a large system. This is OK,
>but fancy channels can save some memory B/W too, which is the
>bottleneck on most full-up server systems. Different approaches.
>The proof of the pudding in this case is that, apparently,
>high-end RISC-based Unix servers *are* getting results
>comparable to the mainframes now. Now, about R-A-S...
Yes, they are, and for the same reasons. Remember that RISC is nothing
more than advertising hype. When I started computing, ALL computers
were RISC - with discrete components, they had to be! Most of the
optimisation techniques used to speed up modern 'RISC' processors were
used on the IBM 360 and 370 ranges, which are often quoted as the
classical example of CISC technology. But this is where we came in :-)
Keep hold of your opinions, and the world will regard you as a wild-eyed
revolutionary, completely up-to-date and hopelessly old-fashioned, with
something like a 15-year cycle.