Sunday, November 19, 2006

Business Rules Forum Benchmarks

Greetings, Programs:

A few things happened during the Q&A part of the BR Forum 2006 presentation on benchmarks. About six minutes into the Q&A part Dr. Forgy said something like this: "I was looking for parallelism in a [inaudible] graphic poblem but I discovered that it had extremely high system demands. For example, I have a parallel application at my office so that I could tell if it is in parallel ... [inaudbile] and it runs Waltz in about a half a second in the serial implementation. When I turn on parallelism, it goes to about 3/4 second." ... "So I've kind of [inaudible] considered [inaudible] Waltz 50 [inaudible] at this point. But there are problems that automatically does [inaudible] where we're going to thave to go to paralleism."

Now, three things to consider: (1) Dr. Forgy is running a rulebase (OPSJ?) in serial mode. (2) Dr. Forgy is running a rulebase(OPSJ?) in parallel mode. (3) either mode of the rulebase runs Waltz 50 is less than a second. One other thing from the conference that is of note from Dr. Forgy was, "There is no known way to solve an NP Complete problem in less than exponential time." Meaning, that there are problems for which ordinary rulebase will not be practical but it will require a parallelized rulebase. Things like Homeland Defense, where there will be hundreds of changes between matches on thousands (or millions) of rules. This will require not only an extremely fast engine but also some sophisticated rule engineering, not the "just throw them in the bucket" kind of business applications that most vendors are doing today in most commercial applications.

One other thing that he noted, "Waltz and Manners are THE benchmarks that are out there. We've looked and looked and looked for others." "There aren't any other benchmarks out there." Well, hopefully, beginning with 2007 we can change that. Personally, Yaakov speaking here, I'm looking forward to doing more than just business applications but we have to realize that it's the business and government applications that pay the bill.

Copyright (c) KBSC


Charles Young said...

The end of using (abusing?) Manners as (and I choose my words carefully here) a general-purpose benchmark for comparison of the performance of different rules engines has my unqualified vote. It is chiefly useful for measuring the performance of negated conjunction at a single node operating under stress, though frankly it would be possible to create a simpler and cleaner test for this single aspect. The increasing dominance of evaluation at a single node as the number of guests is increased is the major contributing factor to the ease with which benchmark implementations can intentionally or inadvertently 'cheat'. Admittedly, the major performance factor for many engines is how well they index their memories, and so in a somewhat coincidental fashion, Manners often reflects this central and important consideration. However, no single benchmark (or even two benchmarks) can adequately provide comprehensive insight into the performance characteristics of an engine. In my opinion, we need a suite of benchmarks that home in on measuring a range of distinct and specific characteristics. This would allow a much more complete picture of the true characteristics of an engine to emerge. Although Peter Lin has disagreed with my stance on Manners, I know he has advocated, and done some preliminary work on designing, just such a suite.

Anonymous said...

Thanks , I've just been searching for info approximately this
subject for ages and yours is the best I've discovered so
far. However, what in regards to the bottom line?
Are you certain about the supply?

My weblog: how to download videos from youtube