|By Warren MacEvoy||
|October 14, 2004 12:00 AM EDT||
/The gains made by better algorithms almost always outstrip the gains from better hardware./
I've frequently seen where algorithm improvements pay by factors of tens to tens of thousands in CPU time. One change I made in a numerical algorithm improved CPU requirements by a factor of 50,000: from weeks on a super-computer to minutes on a workstation.
Any business-savvy engineer knows that algorithm improvements come at a price: the engineer's time. Striking that balance makes software systems move forward rather than staggering to a halt in bloat and dysfunction. It also helps to use people who actually know what they are doing: knowing how to compile code doesn't make you a software engineer any more than knowing how to spell makes you a writer. End of rant.
On to (rant related) business. On most Web sites, think of how many times a data source will be used to retrieve the same data and produce the same content over and over again. Most successful services deliver a highly redundant amount of information to their users. For example, the JDJ website will deliver this (same) content to perhaps a hundred thousand users. If the servers are overtaxed, customers will experience significant delays or malfunctions.
There are several useful solutions to this. Well configured caching proxy servers come to mind, although server-side scripting make this difficult. Buying more hardware will eventually fix the problem, which may be the correct business solution.
But what about asking programmers to be a little more lazy?
For this article I've included the source for the LazyFileOutputStream. It acts just like a regular FileOutputStream except that, if created on a file that already exists, it /reads/ the data from the file instead of writes it. The stream compares what is already in the file with what you are currently writing to it. If at any point it sees there is a difference in the data you are writing this time compared to what is already there, the stream automatically switches to a write-mode that writes over the remainder of the file with the changes.
The upshot is, if your program generates the same output twice, the output file is unmodified the second time (leaving the original modification date). First, by simply changing FileOutputStream to LazyFileOutputStream, any downstream processing can use timestamp information on the files to check if they need to do anything at all. If the timestamp hasn't changed, then neither has the contents.
But wait, there's more! In addition to the standard close(), the LazyFileOutputStream also supports abort(). This method effectively states "I'm done now, leave the rest of the file alone." The remainder of the file will be the same, even without reproducing it. This means that, if you determine at an early point in the processing of the file that it's going to turn out the same, you can simply abort() to leave it alone. Its similar to the idea of not changing the modification dates on files which are rewritten with the same data, but allows for saving CPU time for the current process step as well as downstream processing..
Certain engines produce part of the template before you can conveniently intervene to decide if you really need to regenerate it. By opening up the output as a Lazy file, you can just abort() early and have the old version, with the the old modification time, around for downstream processing.
Okay, rant concluded and point made: CPUs around the planet are spinning through the same data tens of thousands of times producing the same content tens of thousands of times. Instead of buying great big servers to manage this, a smart caching policy based on lazy file writers and some modification time testing could save some sites that same wild-sounding factor of 50,000. Without having to buy 50,000 new servers.
Anecdote # 1. There is a certain technical advantage to this style of writing data as well: most storage devices are easier to read from than write to, adhering to the 80/20 rule: 80% of file access will be reads, 20% will be writes.. The LazyFileOutputStream takes advantage of that for the many files which are simply rewritten with the same content.
Anecdote # 2. There must be a few curled toes out there saying to themselves, "Why not LazyFileWriter?" There are good technical reasons for the OutputStream: the logic of the data written must be checked in its raw /byte/ format for the idea to work correctly, and you can always wrap this in an OutputStreamWriter, followed by a BufferedWriter, which is what I recommend.
Now I'm even done with the anecdotes. Have a nice day.
|Bruce VanOrder 10/19/04 09:25:42 AM EDT|
I remember the first PC I bought for myself ... A CompuAdd 286 with lots of memory - 2 MB RAM and a whopping 40Mb hard drive ! On this gargantuan drive I was able to put everytrhing I needed.... WordPerfect 5.1, TurboPascal 5.5, Lotus 123, dBaseIII+, etc, etc, and a few games .... AND I STILL HAD ROOM !
Now I have a Pentium with 256Mb RAM & a 20 Gb hard drive...
dBaseIII+ could fit on a 1.44Mb 3.5" floppy !
Those were the days my friend, we thought would never end ..
|Warren MacEvoy 10/16/04 12:57:28 AM EDT|
Response to Mark M.
Back in the bad-old-days designers would kick around which sort would be better to use. Now practically all sort problems are best solved with Collections.sort() or using a TreeSet or TreeMap. This is a total win situation: it's faster to write, easier to maintain, and better optimized than any roll-your-own sort. So there's almost no context: the Collections' sort is almost always better.
I'll claim LazyFileOutputStream sits one step lower than this: it's almost never worse, and sometimes better than a plain FileOutputStream. If you are writing small chunks to an unbuffered stream (or flushing() after every character), then the adapter pattern it uses to implement its magic may cost you a little in time (but negligibly compared to other costs related to this approach). There is also a buffer overhead because of the (IMHO silly) decision to leave fundamental memory operations like POSIX memcmp out of the java system libraries. But you're writing to a file, and well, that's just kinda slow.
But what you gain is information. When you're done .isDifferent() will tell you if there was a change without having to keep the old copy around to see if there was a change, and the timestamps will tell you even if downstream processing occurs in some logically distant place, like another process.
So there's very little to lose in almost any situation, and a great deal to gain if:
1. your template processing is file-based.
Without timestamp information, implementing part 2 may have seemed like a waste of time (which it would have been, since every template rebuild would look like it was different), but switching to LazyFileOutputStreams can make it effective.
|Warren MacEvoy 10/15/04 07:07:48 PM EDT|
My apologies, but somehow the wrong link was placed for the source file. The correct adddress to the LazyFileOutputStream which the article refers to is:
You might also note that the class uses abandon() instead of abort(), which is a minor change.
This has nothing to do with
Again, my apologies for any confusion this may have created...
|Mark M 10/15/04 11:47:56 AM EDT|
Response to Warren M.
The key question is not how much more complicated it is to write the class. The key question is what is the context of the problem? Too often generic solutions to problems are presented (even if that is not always the author's intention, these things can be easily mis-interpreted as such) and their validity/necessity almost always depends upon the context of the problem. You yourself emphasize the need for context in your response to Jim M. You have created a useful tool for yourself given the context of the problem you were trying to solve. When the next programmer comes along, the context may be completely different. Often times, many are lead to believe incorrectly in one size fits all philosophies, for instance, it is widely viewed within the industry by working folk like myself that the notions of Bertand Meyer and Kent Beck conflict when in fact they both may be valid solutions under differing contexts. Lack of context is the biggest complaint I have with books on process in this industry. Without it, many arguments are neither valid nor invalid, just ambiguous. There is at least one really bright fella who says a lot about context when he writes. His name is Fred Brooks.
|Warren MacEvoy 10/14/04 10:42:09 PM EDT|
Response to Mark M.
I agree that it usually a waste of time to optimize without profiling to know where your problems are. You must also have a business argument that the problem needs to be solved and that optimizations are the best way to solve it.
It is wrong to think that optimizations must be complicated. There's plenty of code out there that make poor or no use of Collections, which would be faster to write, maintain and execute if better choices were made. Good programmers should know how to use these features to improve turnaround, defects, and efficiency (the rant part of my article).
The purpose of the article is to point another kind of "low hanging fruit" related to file processing. After all, how much more complicated is it to write "LazyFileOutputStream" compared to "FileOutputStream"?
Response to Jim M.
Completely? Substantially. Completely claims they have nothing to learn from each other, yet there are many business problems with a short lifetime and plenty of rustic scientific codes are dutifully solving the problems they were designed to solve twenty and thirty years after they were written.
Again, the optimizations I'm suggesting don't need to be complicated. The LazyFileOutputStream is as simple as the code it replaces. How does that detriment readability or maintainability?
As far as longevity, I like the analogy of building a wall. The last row of the wall (business or scientific) can be very slipshod and the wall will still be a wall. Much software is written with the (sometimes correct) assumption that they will be part of the last row of bricks. But people change their minds, and what was once the last row is not anymore. In the real world, this is why tens of thousands of people die when there is an earthquake in a third-world country.
Should businesses be happy with a software design model analogous to the slums of Mexico City?
Response to Justin S.
Edit one line of an XML configuration file, changing one attribute. Many elements of your design depend on this XML file, but almost none of them depend on this one attribute. Your solution suggests detailed code to see if the attributes each dependency requires has changed, which would be hugely complicated to write and maintain.
Mine asks that you to rebuild the elements that directly depend on the configuration file. If they don't change, then you don't have to propagate updates further. Not a perfect optimization, but a much more practical one.
The LazyFileOutputStream supports your idea if you choose to pursue it. If a template decides it does not need to regenerate a target, it can simply abort() to leave the current contents alone without going to the trouble of regenerating all of it.
|Justin Sadowski 10/14/04 08:27:34 PM EDT|
While I agree with your thoughts about the value of avoiding writing the same data over and over, I have to disagree with your LazyFileOutputStream solution. If you find that you are writing the same data to the same file repeatedly, I would suggest that you improve this by avoiding the rewriting altogether, instead of just making the rewriting more efficient.
For example, perhaps you are writing the same output repeatedly because you are operating on the same input; i.e. the data in a database hasn't changed, or a source XML file hasn't changed. If you can detect your input hasn't been modified, you can avoid writing the output altogether.
I would like to hear more details about the specific situation(s) that you have used LazyFileOutputStream -- I would be interested to hear an example of a situation where my logic above does not apply.
|Jim T. 10/14/04 07:09:02 PM EDT|
Scientific computing and business computing are completely different. In the scientific communality you usually have a very small number of highly skilled people working on a program. That just isn't so in the business world. In the business world I care much more about readability and maintainability rather than speed for 99% of our code. In science, nobody will be using my programs 5 years from now, the data will have all been analyzed and the papers published, in business, the exact same code will be used 5 years from now (or at least it will be for the basis of the code). I believe this is true because it is true of my code from 5 years ago. The physics code is gone/useless and the business code is being resold every day.
|mark mcconkey 10/14/04 05:51:12 PM EDT|
Several years back I began reading Kent Beck's stuff (XP) and it struck a chord for me because many of my experiences were similar. I believe Kent's general notion is something to the effect that one should not optimize up front because its too difficult to predict the future, and the majority of the time you will have made your code unreadable for no reason whatsoever. Of course, any seasoned programmer has experienced enough to have a feel for when big troubles are over the hill and thus that some optimization up front will be needed. I think though, that what is missing in your article is a lack of discussion of context. If I have 3 weeks to finish something that will take 6 and 50 big whigs in a fortune 500 company have goals dependent upon the completion of my software, it doesn't matter how clever I am. It matters how fast I can produce what is needed. On the other hand, the creators of Amazon probably needed to be quite clever in order to deal with the magnitude of hits on their servers. Without context, its sort of useless to talk about optimization.
- Where Are RIA Technologies Headed in 2008?
- The Next Programming Models, RIAs and Composite Applications
- AJAX World RIA Conference & Expo Kicks Off in New York City
- Constructing an Application with Flash Forms from the Ground Up
- Building a Zip Code Proximity Search with ColdFusion
- Personal Branding Checklist
- CFEclipse: The Developer's IDE, Eclipse For ColdFusion
- Has the Technology Bounceback Begun?
- Adobe Flex 2: Advanced DataGrid
- i-Technology Viewpoint: We Need Not More Frameworks, But Better Programmers
- Web Services Using ColdFusion and Apache CXF
- Cloud People: A Who's Who of Cloud Computing