IBM 3090 Model 600J performance from Hercules - Is it possible?

[Sternbach, William]

Hi,

I used to work on an IBM 3090 Model 600J, which at the time, was the
highest capacity mainframe IBM sold and
was usually purchased by the largest corporations that had huge nightly
cycles with thousands
of jobs which had to run concurrently, along with multiple 24 hour / 7
day CICS regions which were accessed by thousands
of users world wide. At the time, that IBM 3090 Model 600J seems like
it had infinite resources.

I'm interested in running the Hercules hardware emulator on a relatively
fast new dual core PC.
I would like to try to potentially achieve the MIPS and throughput
rating of an IBM 3090 Model 600J using
the Hermes hardware emulator. I am not sure if this is possible, but
was wondering if someone on this list
would kindly tell me the performance in terms of MIPS and throughput I
could achieve with the Hercules hardware emulator
on the computer I have (see below), and whether it could ever remotely
approach the performance of a real
IBM 3090 Model 600J.

The PC I have is:
Dell XPS 410 with Intel dual core E6700 CPU running at 2.66 MHZ (Very
high PCMark05 Cpu benchmark score of 6,789),
with 4 GB fast 667 MHZ Ram, and 1 Terabyte of 7200 RPM Western Digital
hard drives running Windows XP Professional SP2.

Bill


[Non-text portions of this message have been removed]

[Greg Smith]

Post by Sternbach, William
I would like to try to potentially achieve the MIPS and throughput
rating of an IBM 3090 Model 600J using the Hermes hardware emulator.

Greek -> Roman:

Hermes ~ Mercury
Herakles ~ Hercules

I think Ivan hit the head on the nail (Ouch!).

We probably execute 50 to 250 host instructions per average emulated
instruction depending on host vs emulated architecture. Some
instructions, like LR, we may only execute 10 host instructions.

Greg

[Sternbach, William]

Hi,

Thank you for your very detailed responses about my question concerning
whether Hercules
on an Intel dual core E6700 could emulate the speed of an IBM 3090 Model
600J.

The link below shows that an IBM 3090 Model 600J ran at 117 MIPS, so I
assume each
of its 6 processors only ran at about (117 / 6 = 19.5 MIPS).

http://www.isham-research.co.uk/mips_chart.html

Jay - This benchmark (http://www.mersenne.org/bench.htm
<http://www.mersenne.org/bench.htm> ) shows the Intel E6700 is about
twice as fast
as the AMD Opterron 275. If we can believe that benchmark, then if
you're getting
"60 MIPS routinely when emulating 4 processors, with real workloads",
maybe Ivan and I
(we both have the Intel E6700) could get "120 MIPS". This assumption
may or may not be correct.
Please correct me if I am wrong.

Also, another important benchmark criteria is throughput.
I remember the throughput of the IBM 3090 Model 600J was very high. It
could handle large numbers
of simultaneously running batch jobs doing huge amounts of IO
simultaneously. I guess its 256 I/O channels
were an important part of this capacity. If we were to compare the I/O
of Hercules running on a desktop PC
with the I/O of a real IBM 3090 Model 600J with its 256 I/O channels,
how much slower would Hercules' I/O be?

Thanks,

Bill








[Non-text portions of this message have been removed]

[Ivan Warren]

Post by Sternbach, William
Hi,
Thank you for your very detailed responses about my question concerning
whether Hercules
on an Intel dual core E6700 could emulate the speed of an IBM 3090 Model
600J.
The link below shows that an IBM 3090 Model 600J ran at 117 MIPS, so I
assume each
of its 6 processors only ran at about (117 / 6 = 19.5 MIPS).

Those figures sound just about right to me !

<snip />

Post by Sternbach, William
Also, another important benchmark criteria is throughput.
I remember the throughput of the IBM 3090 Model 600J was very high. It
could handle large numbers
of simultaneously running batch jobs doing huge amounts of IO
simultaneously. I guess its 256 I/O channels
were an important part of this capacity. If we were to compare the I/O
of Hercules running on a desktop PC
with the I/O of a real IBM 3090 Model 600J with its 256 I/O channels,
how much slower would Hercules' I/O be?
Thanks,
Bill

The problem is not how much slower the I/O could or may be.. (actually,
an individual I/O burst may be faster on hercules than it was on a 3090
600j equipped with 4.5 MBs blue cables hooked up to a 3990)..

The real problem is that any significant I/O may (and will) slow down
the CPU emulation because the CPU execution and I/O tasks are shared
among the physical processors... Let's take an example..

You have 8 physical CPUs.. and 6 emulated CPUs.. Now.. All CPUs are
loaded with some workload and all of a sudden, you have 4 outstanding
I/Os to dasd.. What will happen ? well.. you are now faced with the
situation where you have 6 emulated CPUs and 4 hot I/Os.. Those I/Os are
going to be stealing some duty cycles from at least 2 of the physical
CPUs servicing CPU emulation tasks - hence considerably reducing their
instruction throughput. Depending on thread scheduling, priority, etc..
and the I/O task being performed, 2 of your CPUs may actually grind to a
halt for some portion of time.

However.. If you are very heavily I/O *AND* CPU constrained, and you
find that your DASD I/O is stealing too much CPU cycles then you may
have to resort to non compressed DASDs (because then, the I/O servicing
time will be mostly handled by the onboard DASD controllers..).. The
I/Os may not go much faster, but at least they won't hog the CPU
(remember.. Today's controllers and HBAs *FLY* compared to what we had
available 15-20 years ago.. Attaining 30-45 MBs in linear read/write
isn't uncommon in your everyday desktop - and that's 10 times faster
that you could attain with a single block multiplexor channel !)

Now.. If you are REALLY heavilly I/O loaded (say, had 100+ simultaneous
active I/O on 100+ different channels) - then there is not much way you
can match that (and it's not really an hercules issue, it's your
underlying hardware - unless you have a setup with 10,15 or 20 Fiber
channel HBAs linked to a good number of separate disk subsystems!)

--Ivan