[self-interest] Performance?

[David Ungar]

Thanks, Michael & Jecel.
Have you tried running the benchmark twice?
I don' t think Self does on stack replacement these days.
Hotspot should be better; there have been many more folks working on 
the compilers for longer.
And SPARC should be better than PPC: those wonderful register windows.

- David

On Jul 2, 2004, at 5:15 PM, Michael Latta wrote:

> As Jecel pointed out, we are not trying to find out how "good" self is.
>   The VM technology is clearly newer/better.
>
> It all started when I did some bad math in my head and thought that
> Squeak was running 1000 cycles / bytecode.  The actual
> numbers are much more reasonable (8 / bytecode or so).  Then I asked if
> the same benchmarks could be run in Self to compare
> them on my machine.
>
> Having just got back from JavaOne I have a related question.  Is the
> HotSpot VM faster than Self?  I will probably try recoding it in java
> just
> for curiosity sake.  It sure is hard to look at the success of Java and
> hold out for the Smalltalk/Self language and environment.  It is
> however the things
> that Java left out from these languages and environments that make them
> so much more productive and enjoyable to work in.  Even the best Java
> IDEs
> with some code replacement features, are no where close to the fluidity
> of the "older" environments.  The Java VMs only allow live replacement
> if you are
> not adding/removing methods or fields!  Talk about user visible
> discontinuity.  What we need is a commercially usable VM that runs
> Self/Smalltalk and can
> interoperate with Java (through SOAP?).
>
> Michael
>
>
>
> On Jul 2, 2004, at 3:01 PM, David Ungar wrote:
>
>> What benchmarks are you running?
>> I believe that bytecodes/sec can be a meaningless figure.
>> Remember that in Smalltalk bitblt is one bytecode, and local variable
>> access is one bytecode.
>>
>> When we did our benchmarking work, we took a (then) medium-sized
>> program, Richards,
>> and wrote it in optimized C++, and in several different styles in 
>> Self,
>> varying in amount
>> of object-orientedness.
>>
>> Remember, that in a Self-style VM, sends to self of messages whose
>> method bodies are small are free.
>> Also, custom control structures you build run just as fast as ifTrue:
>> while:, etc.
>> Last time I looked, this was not true in Smalltalk VMs. So, if your
>> benchmark uses a control structure
>> that the ST compiler happens to replace with branch bytecodes, you do
>> OK.
>> But if not, you lose. I don't know if it is still true, but in ST-80,
>> folks used to distort their code
>> to use the "good" control structures.
>>
>> - David
>>
>>
>>
>>
>> On Jul 2, 2004, at 10:28 AM, Jecel Assumpcao Jr wrote:
>>
>>> On Thursday 01 July 2004 23:11, Michael Latta wrote:
>>>> It looks like the compiler on the Mac is not as good as that for the
>>>> Sparc.  On my 1Ghz powerbook I got the following:
>>>>
>>>> 47,058,823 bytecodes/sec and 17,247,597 sends/sec
>>>>
>>>> which is a lot less than the 277Mhz sparc for bytecodes, but
>>>> competitive on sends.  Still not what you want when comparing to
>>>> hardware with 1/4 the clock rate.
>>>
>>> Is this a G4? I'll see what numbers I get on 1GHz G4 eMac the next
>>> time
>>> I get to use that machine. I suppose you are using the latest version
>>> -
>>> Self on the Mac didn't have the SIC (simple inlining compiler) before
>>> 4.2 and didn't even have PICs (polymorphic inline caches) before
>>> 4.1.6.
>>>
>>> <http://research.sun.com/self/release_4.2/Self-4.2.1/Self-4.2.1-
>>> Release-Notes.pdf>
>>>
>>> It seems that the compilers have been improved for the Sparc. Here 
>>> are
>>> more numbers on that same machine:
>>>
>>> Self 4.0 =>  60,150,375 bytecodes/sec; 5,845,070 sends/sec
>>> Self 4.1 =>  78,527,607 bytecodes/sec; 15,404,782 sends/sec
>>>
>>> You might also want to run the benchmarks several times. Here are the
>>> first ten runs in Self 4.1.1 on the Ultra:
>>>
>>>   75,029,308 bytecodes/sec; 16,746,759 sends/sec
>>>   74,116,965 bytecodes/sec; 16,915,608 sends/sec
>>>   69,678,824 bytecodes/sec; 16,492,340 sends/sec
>>>   72,480,181 bytecodes/sec; 16,946,674 sends/sec
>>>   97,116,843 bytecodes/sec; 16,962,250 sends/sec
>>>  100,078,186 bytecodes/sec; 16,993,488 sends/sec
>>>  101,910,858 bytecodes/sec; 16,962,250 sends/sec
>>>  101,185,770 bytecodes/sec; 16,977,855 sends/sec
>>>  101,105,845 bytecodes/sec; 17,024,842 sends/sec
>>>   85,106,382 bytecodes/sec; 17,087,897 sends/sec
>>>
>>> Please note that these are bad benchmarks for Squeak, and even worse
>>> for
>>> Self (the moral equivalent of BogoMIPS in Linux). For a serious
>>> comparison we should be using DeltaBlue, Richards and so on.
>>>
>>>> Now I need to run the same thing on Squeak to compare on the same
>>>> hardware.  Could you point me at the squeak expression?
>>>
>>>    0 tinyBenchmarks
>>>
>>> -- Jecel
>>>
>>>
>>>
>>> Yahoo! Groups Links
>>>
>>>
>>>
>>>
>>>
>>
>>
>>
>>
>> Yahoo! Groups Links
>>
>>
>>
>>
>>
>
>
>
>
> Yahoo! Groups Links
>
>
>
>
>