home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.lang.smalltalk
- Path: sparky!uunet!charon.amdahl.com!amdahl!rtech!pacbell.com!ames!saimiri.primate.wisc.edu!zaphod.mps.ohio-state.edu!cs.utexas.edu!news.uta.edu!utafll!bruce
- From: bruce@utafll.uta.edu (Bruce Samuelson)
- Subject: for Claus Gittinger
- Message-ID: <1993Jan26.191446.16764@utagraph.uta.edu>
- Sender: news@utagraph.uta.edu (USENET News System)
- Nntp-Posting-Host: utafll.uta.edu
- Organization: University of Texas at Arlington
- Date: Tue, 26 Jan 1993 19:14:46 GMT
- Lines: 65
-
- Claus, here is email I tried to send you but my mailer returned with
- host unknown.
-
- Date: Tue, 26 Jan 93 13:00:18 CST
- From: bruce (Bruce Samuelson)
- To: gitting@slsvftt.us-es.sel.de
- Subject: your comments on slopstones and smopstones
-
- Thanks for your comments, which are all valid. I have received similar
- comments privately and publicly, and I agree that Slopstones in particular has
- shortcomings. I tried developing several earlier versions prior to the one I
- published. All the earlier ones were either flawed worse, or made ST/V-DOS
- crash when I ran them.
-
- Slopstones is intended to measure strictly low level operations and will not
- necessarily reflect performance in an actual application. I wanted to measure
- seven common operations and get the fastest time I could for each. I agree
- that a compiler could theoretically compile away some of them, and in fact, if
- you had seen a posting a couple weeks ago by Urs Hoelzle, Self did in fact do
- just that. You are correct, though, that theoretically they cannot be
- optimized away, because one could always redefine integer addition, for
- example, to not only add two integers, but to do something else such as keep
- track of how many times integer addition has been invoked.
-
- Fractonaccis: yes, fractions in themselves are not that interesting, but the
- machinery needed to add/subtract them does invoke quite a lot of integer
- arithmetic, which *is* interesting. However, I did not choose fractions
- because they were interesting, but because fibonaccis do not produce usable
- runs. Either the numbers grow larger than 16-bit integer limits on ST/V-DOS
- (16363) or the run time is too small to be reliably measured. Regarding
- breaking the fractonacci test into recursion plus something else, that would
- have been a possibility. I'll comment on this more later.
-
- [] value: I agree that passing arguments and referencing method variables
- would have also been interesting tests. I limited myself severely as to the
- number of tests I did because ST/V-DOS was too crash-prone whenever I got too
- ambitious. If I had not been constrained to 16-bit integers and ST/V-DOS
- compatibility, I could have done more that was interesting. The reason I broke
- the tests up into a 'setup' method and 'execute' method was that the ST/V-DOS
- compiler could not handle it when it was combined into one large method.
-
- Your simple recursion test: this was almost identical to the one posted by
- Marten Feldtmann a few weeks ago. It turned out that ST/V-PM outperformed
- ST80-Win on his benchmark. I wrote Slopstone and Smopstone to see whether it
- would also beat ST80 on a wider range of tests. In general, I don't think ST80
- is very well optimized for recursive calls.
-
- Your sieve test: I wish I had this when I wrote Smopstones. I probably would
- have included it.
-
- Conclusion: Thank you for your constructive criticisms. I would incorporate
- some of your suggestions into a revised form of Slopstone and Smopstone, but I
- don't have the time. They are already set in concrete, and quite a few results
- have already been tabulated for them. However, perhaps in the future, you or I
- or someone else may offer a new set of benchmarks that are more theoretically
- sound and well balanced. Thanks again for your comments.
-
- I plan to write a short article for The Smalltalk Report on benchmarking, and
- will cite your concerns there, along with similar concerns from others. I'll
- keep your name anonymous unless you tell me otherwise.
- --
- **********************************************************
- * Bruce Samuelson Department of Linguistics *
- * bruce@ling.uta.edu University of Texas at Arlington *
- **********************************************************
-