[Info-vax] Emulation Performance (was: Re: Unix on A DEC Vax?)

[Stephen Hoffman]

TL;DR: very little performance data is around; I haven't seen any 
published list.  I wouldn't expect the performance of even the best 
available emulator to excel.

On 2013-01-17 08:18:48 +0000, John Wallace said:

> On Jan 17, 1:38 am, Stephen Hoffman <seaoh... at hoffmanlabs.invalid>
> wrote:
>> On 2013-01-16 23:13:14 +0000, ChrisQ said:
>> 
>>> With any vax though, don't expect 3Ghz Xeon performance :-)...
>> 
>> Or with a 3 GHz Xeon and a software emulator, expect to get VAX
>> performance.  :-)
> 
> 
> What kind of performance is reasonable to expect from emulated
> systems?

Hopefully enough performance to get the application(s) ported to newer 
hardware, or replaced.

Put another way, emulation is a means to a goal, whether that's 
admitted or not.

That's because of just what you're looking at; the performance.  That, 
and the eventual pain in the rump of maintaining the whole stack.

For an AlphaServer 2100 5/250, I'd look to transfer the licenses over 
to a used EV6 or EV7-class Alpha, and move on.  Or port to Itanium, and 
move on.  Or some combination of these, and possibly emulation, and a 
port of either the apps or (more commonly) the data.


> I know that's a "how long is your piece of string" question,
> but is it answered anywhere,objectively or anecdotally?

I don't have any good data, and I've not seen any collected list.

It would make sense to run some number of UETP passes on some random 
boxes running various emulators, and scrounge up some numbers.  (Maybe 
with ALL, 1 pass, 4 users?  At the UETP defaults, at least to start?  
Other suggestions?)

With one case I'm aware of, a half-GHz x86 box (obviously not the 
fastest hardware around) was getting low- to mid-range VAX speeds.  The 
most recent simh boot I did was pretty speedy, but (as is usual) hogged 
a whole core, and spun up the fans.  (Yes, I'm aware of the VAX MP 
multiprocessing simh fork/port/variant.  
<http://www.oboguev.net/vax_mp/>  No, I haven't tried that yet.)

> The test workload on the emulator was three times slower (ie runtime
> three times longer) [than the original AlphaServer 2100 5/250 -Hoff], 
> which was a bit of a surprise.
> 
> Is that kind of performance degradation (vs the same workload on
> antique native hardware) to be expected?

For instruction emulation, that's surprisingly good.

For a JITting emulator, that seems decent performance, but that also 
depends on the app, and on what the app and the OS are doing.

FX!32 was probably one of the most effective designs here, as that 
could call through to the "real" Windows routines.


> Does performance vary widely
> with different emulators?

Yes.  Performance does vary.  Probably most centrally around support 
for JITting, but also the instruction (in)efficiencies.

> Does the host environment (e.g. VMware vs
> native) matter much?

Donno.  Figure out how well a compute-bound OS does in your chosen 
virtual machine.

Emulation has a very long history in computing.  As a stopgap measure.  
As a way-place on the path.  As a tool to contend with planning gaps or 
budgetary shortfalls.  Until the old software ages out, or the folks 
that want and use it age out.  Or until management decides to burn down 
the parts of the organization involved and either rebuild it or sell 
off the assets, or whatever the current euphemism is.

While these are tools that work and that can be useful for some cases, 
I'm generally not a proponent of either translation, or of emulation.

And yes, Java and the other languages that are hosted on the JVM are 
the exception to this.  Performance isn't grand, but does pretty well 
on most current boxes.  But the JVM stack has had a whole lot more 
development effort than has any of the JITting emulators.


-- 
Pure Personal Opinion | HoffmanLabs LLC