Re: performance coding project? (was: Re: When to cache)
On Sat, 26 Jan 2002, Stas Bekman wrote: [...] It's much better to build your system, profile it, and fix the bottlenecks. The most effective changes are almost never simple coding changes like the one you showed, but rather large things like using qmail-inject instead of SMTP, caching a slow database query or method call, or changing your architecture to reduce the number of network accesses or inter-process communications. It all depends on what kind of application do you have. If you code is CPU-bound these seemingly insignificant optimizations can have a very significant influence on the overall service performance. Of course if you app, is IO-bound or depends with some external component, than your argumentation applies. Eh, any real system will be a combination. Sure; when everything works then it's worth finding the CPU intensive places and fix them up, but for the most part the system design is far far more important than any code optimiziation you can ever do. My usual rhetorics: Your average code optimization will gain you at most a few percent performance gain. A better design can often make things 10 times faster and use only a fraction of your memory. On the other hand how often do you get a chance to profile your code and see how to improve its speed in the real world. Managers never plan for debugging period, not talking about optimizations periods. And while premature optimizations are usually evil, as they will bait you later, knowing the differences between coding styles does help in a long run and I don't consider these as premature optimizations. If you don't waste time profiling every little snippet of code you might have more time to fix the real bottlenecks in the end. ;-) [...] All I want to say is that there is no one-fits-all solution in Perl, because of TIMTOWTDI, so you can learn a lot from running benchmarks and picking your favorite coding style and change it as the language evolves. But you shouldn't blindly apply the outcomes of the benchmarks without running your own benchmarks. Amen. (And don't get me wrong; I think a repository of information about the nitty gritty optimization things would be great - I just find it to be bad advice to not tell people to do the proper design first). - ask -- ask bjoern hansen, http://ask.netcetera.dk/ !try; do(); more than a billion impressions per week, http://valueclick.com
Re: performance coding project? (was: Re: When to cache)
On Sat, 26 Jan 2002, Perrin Harkins wrote: It all depends on what kind of application do you have. If you code is CPU-bound these seemingly insignificant optimizations can have a very significant influence on the overall service performance. Do such beasts really exist? I mean, I guess they must, but I've never seen a mod_perl application that was CPU-bound. They always seem to be constrained by database speed and memory. At ValueClick we only have a few BerkeleyDBs that are in the request loop for 99% of the traffic; everything else is in fairly efficient in-memory data structures. So there we do of course care about the tiny small optimiziations because there's a direct correlation between saved CPU cycles and request capacity. However, it's only that way because we made a good design for the application in the first place. :-) (And for all the other code we rarely care about using a few more CPU cycles if it is easier/cleaner/more flexible/comes to mind first). Who cares if the perl code gets ready to wait for the database a few milliseconds faster? :-) - ask -- ask bjoern hansen, http://ask.netcetera.dk/ !try; do(); more than a billion impressions per week, http://valueclick.com
Re: performance coding project? (was: Re: When to cache)
One memory speed saving is using global VARS, I know it is not recomended practice, but if from the begining of the project u set a convention what are the names of global vars it is ok..F.e. I'm using in all DB pages at the begining : our $dbh = dbConnect() or dbiError(); See I know (i'm sure) that when I use DB I will always initialize the var. One other example is (ASP.pm): our $userID = $$Session{userID}; my $something = $$Request{Params}{something} This is not saving me memory, but shorten my typewriting especialy if it is used frequently or if I need to change FORM-param from something to anything..etc.. I think in mod_perl world we are about to search MEMORY optimisation RATHER speed optimisation... :) raptor [EMAIL PROTECTED]
Re: performance coding project? (was: Re: When to cache)
Hi all, Stas has a point. Perl makes it very easy to do silly things. This is what I was doing last week: if( m/\b$Needle\b/ ) {...} Eight hours. (Silly:) if( index($Haystack,$Needle) m/\b$Needle\b/ ) {...} Twelve minutes. 73, Ged.
Re: performance coding project? (was: Re: When to cache)
It all depends on what kind of application do you have. If you code is CPU-bound these seemingly insignificant optimizations can have a very significant influence on the overall service performance. Do such beasts really exist? I mean, I guess they must, but I've never seen a mod_perl application that was CPU-bound. They always seem to be constrained by database speed and memory. On the other hand how often do you get a chance to profile your code and see how to improve its speed in the real world. Managers never plan for debugging period, not talking about optimizations periods. If you plan a good architecture that avoids the truly slow stuff (disk/network access) as much as possible, your application is usually fast enough without spending much time on optimization (except maybe some database tuning). At my last couple of jobs we actually did have load testing and optimization as part of the development plan, but that's because we knew we'd be getting pretty high levels of traffic. Most people don't need to tune very much if they have a good architecture, and it's enough for them to fix problems as they become visible. Back to your idea: you're obviously interested in the low-level optimization stuff, so of course you should go ahead with it. I don't think it needs to be a separate project, but improvements to the performance section of the guide are always a good idea. I know that I have taken all of the DBI performance tips to heart and found them very useful. I'm more interested in writing about higher level performance issues (efficient shared data, config tuning, caching), so I'll continue to work on those things. I'm submitting a proposal for a talk on data sharing techniques at this year's Perl Conference, so hopefully I can contribute that to the guide after I finish it. - Perrin
Re: performance coding project? (was: Re: When to cache)
On Sat, 26 Jan 2002, Perrin Harkins wrote: It all depends on what kind of application do you have. If you code is CPU-bound these seemingly insignificant optimizations can have a very significant influence on the overall service performance. Do such beasts really exist? I mean, I guess they must, but I've never seen a mod_perl application that was CPU-bound. They always seem to be constrained by database speed and memory. I've seen one. However, it was much like a normal performance problem - the issue was with one loop which ran one line which was quite pathological. Replacing loop with an s///eg construct eliminated the problem; there was no need for seemlingly insignificant optimizations. (Actually, the problem was *created* by premature optimization - the coder had utilized code that was more efficient than s/// in one special case, to handle a vastly different instance.) However, there could conceivably be code which was more of a performance issue, especially when the mod_perl utilizes a very successful cache on a high traffic site. On the other hand how often do you get a chance to profile your code and see how to improve its speed in the real world. Managers never plan for debugging period, not talking about optimizations periods. Unless there's already a problem, and you have a good manager. We've had a couple of instances where we were given time (on the schedule, before the release) to improve speed after a release. It's quite rare, though, and I've never seen it for a mod_perl project. Ed
Re: performance coding project? (was: Re: When to cache)
On Sat, 26 Jan 2002, Perrin Harkins wrote: It all depends on what kind of application do you have. If you code is CPU-bound these seemingly insignificant optimizations can have a very significant influence on the overall service performance. Do such beasts really exist? I mean, I guess they must, but I've never seen a mod_perl application that was CPU-bound. They always seem to be constrained by database speed and memory. Think search engines. Once you've figured out how to get your search database to fit in memory (or devised a cachin strategy to get the important parts there) you're essentially looking at a CPU-bound problem. These days the best solution is probably some judicious use of Inline::C. Back when I last tackled the problem I had to hike up mount XS to find my grail... -sam
Re: performance coding project? (was: Re: When to cache)
On Saturday 26 January 2002 03:40 pm, Sam Tregar wrote: Think search engines. Once you've figured out how to get your search database to fit in memory (or devised a cachin strategy to get the important parts there) you're essentially looking at a CPU-bound problem. These days the best solution is probably some judicious use of Inline::C. Back when I last tackled the problem I had to hike up mount XS to find my grail... I agree. There are some situations that are just too complex for a DBMS to handle directly, at least in any sort of efficient fashion. However, depending on the load in those cases, Perrin's solution for eToys is probably a good approach (i.e. custom search software written in C/C++). -- Milo Hyson CyberLife Labs, LLC
Re: performance coding project? (was: Re: When to cache)
Perrin Harkins wrote: Back to your idea: you're obviously interested in the low-level optimization stuff, so of course you should go ahead with it. I don't think it needs to be a separate project, but improvements to the performance section of the guide are always a good idea. It has to be a run-able code, so people can verify the facts which may change with different OS/versions of Perl. e.g. Joe says that $r-args is slower then Apache::Request-param, I saw the opposite. Having these as a run-able bits, is much nicer. I know that I have taken all of the DBI performance tips to heart and found them very useful. :) That's mostly JWB's work I think. I'm more interested in writing about higher level performance issues (efficient shared data, config tuning, caching), so I'll continue to work on those things. I'm submitting a proposal for a talk on data sharing techniques at this year's Perl Conference, so hopefully I can contribute that to the guide after I finish it. Go Perrin! _ Stas Bekman JAm_pH -- Just Another mod_perl Hacker http://stason.org/ mod_perl Guide http://perl.apache.org/guide mailto:[EMAIL PROTECTED] http://ticketmaster.com http://apacheweek.com http://singlesheaven.com http://perl.apache.org http://perlmonth.com/
Re: performance coding project? (was: Re: When to cache)
Ah yes, but don't forget that to get this speed, you are sacrificing memory. You now have another locally scoped variable for perl to keep track of, which increases memory usage and general overhead (allocation and garbage collection). Now, those, too, are insignificant with one use, but the significance will probably rise with the speed gain as you use these techniques more often... Issac Stas Bekman wrote: Rob Nagler wrote: Perrin Harkins writes: Here's a fun example of a design flaw. It is a performance test sent to another list. The author happened to work for one of our competitors. :-) That may well be the problem. Building giant strings using .= can be incredibly slow; Perl has to reallocate and copy the string for each append operation. Performance would likely improve in most situations if an array were used as a buffer, instead. Push new strings onto the array instead of appending them to a string. #!/usr/bin/perl -w ### Append.bench ### use Benchmark; sub R () { 50 } sub Q () { 100 } @array = ( x R) x Q; sub Append { my $str = ; map { $str .= $_ } @array; } sub Push { my @temp; map { push @temp, $_ } @array; my $str = join , @temp; } timethese($ARGV[0], { append = \Append, push = \Push }); Such a simple piece of code, yet the conclusion is incorrect. The problem is in the use of map instead of foreach for the performance test iterations. The result of Append is an array of whose length is Q and whose elements grow from R to R * Q. Change the map to a foreach and you'll see that push/join is much slower than .=. Return a string reference from Append. It saves a copy. If this is the page, you'll see a significant improvement in performance. Interestingly, this couldn't be the problem, because the hypothesis is incorrect. The incorrect test just validated something that was faulty to begin with. This brings up you can't talk about it unless you can measure it. Use a profiler on the actual code. Add performance stats in your code. For example, we encapsulate all DBI accesses and accumulate the time spent in DBI on any request. We also track the time we spend processing the entire request. While we are at this topic, I want to suggest a new project. I was planning to start working on it long time ago, but other things always took over. The perl.apache.org/guide/performance.html and a whole bunch of performance chaptes in the upcoming modperl book have a lot of benchmarks, comparing various coding techniques. Such as the example you've provided. The benchmarks are doing both pure Perl and mod_perl specific code (which requires running Apache, a perfect job for the new Apache::Test framework.) Now throw in the various techniques from 'Effective Perl' book and voila you have a great project to learn from. Also remember that on varous platforms and various Perl versions the benchmark results will differ, sometimes very significantly. I even have a name for the project: Speedy Code Habits :) The point is that I want to develop a coding style which tries hard to do early premature optimizations. Let me give you an example of what I mean. Tell me what's faster: if (ref $b eq 'ARRAY'){ $a = 1; } elsif (ref $b eq 'HASH'){ $a = 1; } or: my $ref = ref $b; if ($ref eq 'ARRAY'){ $a = 1; } elsif ($ref eq 'HASH'){ $a = 1; } Sure, the win can be very little, but it ads up as your code base's size grows. Give you a similar example: if ($a-lookup eq 'ARRAY'){ $a = 1; } elsif ($a-lookup eq 'HASH'){ $a = 1; } or my $lookup = $a-lookup; if ($lookup eq 'ARRAY'){ $a = 1; } elsif ($lookup eq 'HASH'){ $a = 1; } now throw in sub attributes and re-run the test again. add examples of map vs for. add examples of method lookup vs. procedures add examples of concat vs. list vs. other stuff from the guide. mod_perl specific examples from the guide/book ($r-args vs Apache::Request::param, etc) If you understand where I try to take you, help me to pull this project off and I think in a long run we can benefit a lot. This goes along with the Apache::Benchmark project I think (which is yet another thing I want to start...), probably could have these two ideas put together. _
Re: performance coding project? (was: Re: When to cache)
Issac Goldstand wrote: Ah yes, but don't forget that to get this speed, you are sacrificing memory. You now have another locally scoped variable for perl to keep track of, which increases memory usage and general overhead (allocation and garbage collection). Now, those, too, are insignificant with one use, but the significance will probably rise with the speed gain as you use these techniques more often... Yes, I know. But from the benchmark you can probably have an idea whether the 'caching' is worth the speedup (given that the benchmark is similar to your case). For example it depends on how many times you need to use the cache. And how big is the value. e.g. may be caching $foo-bar doesn't worth it, but what about $foo-bar-baz? or if you have a deeply nested hash and you need to work with only a part of subtree, do you grab a reference to this sub-tree node and work it or do you keep on dereferencing all the way up to the root on every call? But personally I still didn't decide which one is better and every time I'm in a similar situation, I'm never sure which way to take, to cache or not to cache. But that's the cool thing about Perl, it keeps you on your toes all the time (if you want to :). BTW, if somebody has interesting reasonings for using one technique versus the other performance-wise (speed+memory), please share them. This project's idea is to give stright numbers for some definitely bad coding practices (e.g. map() in the void context), and things which vary a lot depending on the context, but are interesting to think about (e.g. the last example of caching the result of ref() or a method call) _ Stas Bekman JAm_pH -- Just Another mod_perl Hacker http://stason.org/ mod_perl Guide http://perl.apache.org/guide mailto:[EMAIL PROTECTED] http://ticketmaster.com http://apacheweek.com http://singlesheaven.com http://perl.apache.org http://perlmonth.com/
Re: performance coding project? (was: Re: When to cache)
This project's idea is to give stright numbers for some definitely bad coding practices (e.g. map() in the void context), and things which vary a lot depending on the context, but are interesting to think about (e.g. the last example of caching the result of ref() or a method call) I think this would be handy. I spend a fair bit of time wondering/testing myself. Would be nice to have a repository of the tradeoffs. OTOH, I spend too much time mulling over unimportant performance optimizations. The foreach/map comparison is a good example of this. It only starts to matter (read milliseconds) at the +100KB and up range, I find. If a site is returning 100KB pages for typical responses, it has a problem at a completely different level than map vs foreach. Rob Pre-optimization is the root of all evil -- C.A.R. Hoare
Re: performance coding project? (was: Re: When to cache)
The point is that I want to develop a coding style which tries hard to do early premature optimizations. We've talked about this kind of thing before. My opinion is still the same as it was: low-level speed optimization before you have a working system is a waste of your time. It's much better to build your system, profile it, and fix the bottlenecks. The most effective changes are almost never simple coding changes like the one you showed, but rather large things like using qmail-inject instead of SMTP, caching a slow database query or method call, or changing your architecture to reduce the number of network accesses or inter-process communications. The exception to this rule is that I do advocate thinking about memory usage from the beginning. There are no good tools for profiling memory used by Perl, so you can't easily find the offenders later on. Being careful about passing references, slurping files, etc. pays off in better scalability later. - Perrin
Re: performance coding project? (was: Re: When to cache)
On Fri, 2002-01-25 at 09:08, Perrin Harkins wrote: snip / It's much better to build your system, profile it, and fix the bottlenecks. The most effective changes are almost never simple coding changes like the one you showed, but rather large things like using qmail-inject instead of SMTP, caching a slow database query or method call, or changing your architecture to reduce the number of network accesses or inter-process communications. qmail-inject? I've just been using sendmail or, preferentially, Net::SMTP. Isn't using a system call more expensive? If not, how does qmail-inject work? Thanks, David -- David Wheeler AIM: dwTheory [EMAIL PROTECTED] ICQ: 15726394 Yahoo!: dew7e Jabber: [EMAIL PROTECTED]
Re: performance coding project? (was: Re: When to cache)
On 25 Jan 2002, David Wheeler wrote: On Fri, 2002-01-25 at 09:08, Perrin Harkins wrote: snip / It's much better to build your system, profile it, and fix the bottlenecks. The most effective changes are almost never simple coding changes like the one you showed, but rather large things like using qmail-inject instead of SMTP, caching a slow database query or method call, or changing your architecture to reduce the number of network accesses or inter-process communications. qmail-inject? I've just been using sendmail or, preferentially, Net::SMTP. Isn't using a system call more expensive? If not, how does qmail-inject work? With qmail, SMTP generally uses inetd, which is slow, or daemontools, which is faster, but still slow, and more importantly, it anyway goes: perl - SMTP - inetd - qmail-smtpd - qmail-inject. So with going direct to qmail-inject, your email skips out a boat load of processing and goes direct into the queue. Of course none of this is relevant if you're not using qmail ;-) -- !-- Matt -- :-Get a smart net/:-
Re: performance coding project? (was: Re: When to cache)
On Fri, 25 Jan 2002 21:15:54 + (GMT) Matt Sergeant [EMAIL PROTECTED] wrote: With qmail, SMTP generally uses inetd, which is slow, or daemontools, which is faster, but still slow, and more importantly, it anyway goes: perl - SMTP - inetd - qmail-smtpd - qmail-inject. So with going direct to qmail-inject, your email skips out a boat load of processing and goes direct into the queue. Of course none of this is relevant if you're not using qmail ;-) Yet another solution: use Mail::QmailQueue, directly http://search.cpan.org/search?dist=Mail-QmailQueue -- Tatsuhiko Miyagawa [EMAIL PROTECTED]
Re: performance coding project? (was: Re: When to cache)
On Fri, 2002-01-25 at 13:15, Matt Sergeant wrote: With qmail, SMTP generally uses inetd, which is slow, or daemontools, which is faster, but still slow, and more importantly, it anyway goes: perl - SMTP - inetd - qmail-smtpd - qmail-inject. So with going direct to qmail-inject, your email skips out a boat load of processing and goes direct into the queue. Okay, that makes sense. In my activitymail CVS script I just used sendmail. http://www.cpan.org/authors/id/D/DW/DWHEELER/activitymail-0.987 But it looks like this might be more efficient, if qmail happens to be installed (not sure on SourceForge's servers). Of course none of this is relevant if you're not using qmail ;-) Yes, and in Bricolage, I used Net::SMTP to keep it as platform-independent as possible. It should work on Windows, even! Besides, all mail gets sent during the Apache cleanup phase, so there should be no noticeable delay for users. David -- David Wheeler AIM: dwTheory [EMAIL PROTECTED] ICQ: 15726394 Yahoo!: dew7e Jabber: [EMAIL PROTECTED]
Re: performance coding project? (was: Re: When to cache)
Stas Bekman [EMAIL PROTECTED] writes: I even have a name for the project: Speedy Code Habits :) The point is that I want to develop a coding style which tries hard to do early premature optimizations. I disagree with the POV you seem to be taking wrt write-time optimizations. IMO, there are precious few situations where writing Perl in some prescribed style will lead to the fastest code. What's best for one code segment is often a mediocre (or even stupid) choice for another. And there's often no a-priori way to predict this without being intimate with many dirty aspects of perl's innards. I'm not at all against divining some abstract _principles_ for accelerating a given solution to a problem, but trying to develop a Speedy Style is IMO folly. My best and most universal advice would be to learn XS (or better Inline) and use a language that was _designed_ for writing finely-tuned sections of code. But that's in the post-working-prototype stage, *not* before. [...] mod_perl specific examples from the guide/book ($r-args vs Apache::Request::param, etc) Well, I've complained about that one before, and since the guide's text hasn't changed yet I'll try saying it again: Apache::Request::param() is FASTER THAN Apache::args(), and unless someone wants to rewrite args() IN C, it is likely to remain that way. PERIOD. Of course, if you are satisfied using Apache::args, than it would be silly to change styles. YMMV -- Joe Schaefer
Re: performance coding project? (was: Re: When to cache)
Perrin Harkins wrote: The point is that I want to develop a coding style which tries hard to do early premature optimizations. We've talked about this kind of thing before. My opinion is still the same as it was: low-level speed optimization before you have a working system is a waste of your time. It's much better to build your system, profile it, and fix the bottlenecks. The most effective changes are almost never simple coding changes like the one you showed, but rather large things like using qmail-inject instead of SMTP, caching a slow database query or method call, or changing your architecture to reduce the number of network accesses or inter-process communications. It all depends on what kind of application do you have. If you code is CPU-bound these seemingly insignificant optimizations can have a very significant influence on the overall service performance. Of course if you app, is IO-bound or depends with some external component, than your argumentation applies. On the other hand how often do you get a chance to profile your code and see how to improve its speed in the real world. Managers never plan for debugging period, not talking about optimizations periods. And while premature optimizations are usually evil, as they will bait you later, knowing the differences between coding styles does help in a long run and I don't consider these as premature optimizations. Definitely this discussion has no end. Everybody is right in their particular project, since there are no two projects which are the same. All I want to say is that there is no one-fits-all solution in Perl, because of TIMTOWTDI, so you can learn a lot from running benchmarks and picking your favorite coding style and change it as the language evolves. But you shouldn't blindly apply the outcomes of the benchmarks without running your own benchmarks. _ Stas Bekman JAm_pH -- Just Another mod_perl Hacker http://stason.org/ mod_perl Guide http://perl.apache.org/guide mailto:[EMAIL PROTECTED] http://ticketmaster.com http://apacheweek.com http://singlesheaven.com http://perl.apache.org http://perlmonth.com/
Re: When to cache
1) The old cache entry is overwritten with the new. 2) The old cache entry is expired, thus forcing a database hit (and subsequent cache load) on the next request. 3) Cache only stuff which doesn't expire (except on server restarts). We don't cache any mutable data, and there are no sessions. We let the database do the caching. We use Oracle, which has a pretty good cache. We do cache some stuff that doesn't change, e.g. default permissions, and we release weekly, which involves a server restart and a refresh of the cache. If you hit http://www.bivio.com , you'll get a page back in under 300ms. There are probably 10 database queries involved if you are logged in. This page is complex, but far from our most complex. For example, this page http://www.bivio.com/demo_club/accounting/investments sums up all the holdings of a portfolio from the individual transactions (buys, sells, splits, etc.). It also comes back in under 300ms. Sorry if this wasn't the answer you were looking for. :) Rob
Re: When to cache
I'm interested to know what the opinions are of those on this list with regards to caching objects during database write operations. I've encountered different views and I'm not really sure what the best approach is. I described some of my views on this in the article on the eToys design, which is archived at perl.com. Take a typical caching scenario: Data/objects are locally stored upon loading from a database to improve performance for subsequent requests. But when those objects change, what's the best method for refreshing the cache? There are two possible approaches (maybe more?): 1) The old cache entry is overwritten with the new. 2) The old cache entry is expired, thus forcing a database hit (and subsequent cache load) on the next request. The first approach would tend to yield better performance. However there's no guarantee the data will ever be read. The cache could end up with a large amount of data that's never referenced. The second approach would probably allow for a smaller cache by ensuring that data is only cached on reads. There are actually thousands of variations on caching. In this case you seem to be asking about one specific aspect: what to cache. Another important question is how to ensure cache consistency. The approach you choose depends on frequency of updates, single server vs. cluster, etc. There's a simple answer for what to cache: as much as you can, until you hit some kind of limit or performance is good enough. Sooner or later you will hit the point where the tradeoff in storage or in time spent ensuring cache consistency will force you to limit your cache. People usually use something like a dbm or Cache::Cache to implement mod_perl caches, since then you get to share the cache between processes. Storing the cache on disk means your storage is nearly unlimited, so we'll ignore that aspect for now. There's a lot of academic research about deciding what to cache in web proxy servers based on a limited amount of space which you can look at if you have space limitations. Lots of stuff on LRU, LFU, and other popular cache expiration algorithms. The limit you are more likely to hit is that it will start to take too long to populate the cache with everything. Here's an example from eToys: We used to generate most of the site as static files by grinding through all the products in the database and running the data through a templating system. This is a form of caching, and it gave great performance. One day we had to add a large number of products that more than doubled the size of our database. The time to generate all of them became prohibitive in that our content editors wanted updates to happen within a certain number of hours but it was taking longer than that number of hours to generate all the static files. To fix this, we moved to not generating anything until it was requested. We would fetch the data the first time it was asked for, and then cache it for future requests. (I think this corresponds to your option 2.) Of course then you have to decide on a cache consistency approach for keeping that data fresh. We used a simple TTL approach because it was fast and easy to implement (good enough). This is just scratching the surface of caching. If you want to learn more, I would suggest some introductory reading. You can find lots of general ideas about caching by searching Google for things like cache consistency. There are also a couple of good articles on the subject that I've read recently. Randal has an article that shows an implementation of what I usually call lazy reloading: http://www.linux-mag.com/2001-01/perl_01.html There's one about cache consistency on O'Reilly's onjava.com, but all the examples are in Java: http://www.onjava.com/pub/a/onjava/2002/01/09/dataexp1.html Also, in reference to Rob Nagler's post, it's obviously better to be in a position where you don't need to cache to improve performance. Caching adds a lot of complexity and causes problems that are hard to explain to non-technical people. However, for many of us caching is a necessity for decent performance. - Perrin
Re: When to cache
Perrin Harkins writes: To fix this, we moved to not generating anything until it was requested. We would fetch the data the first time it was asked for, and then cache it for future requests. (I think this corresponds to your option 2.) Of course then you have to decide on a cache consistency approach for keeping that data fresh. We used a simple TTL approach because it was fast and easy to implement (good enough). I'd be curious to know the cache hit stats. BTW, this case seems to be an example of immutable data, which is definitely worth caching if performance dictates. However, for many of us caching is a necessity for decent performance. I agree with latter clause, but take issue with the former. Typical sites get a few hits a second at peak times. If a site isn't returning typical pages in under a second using mod_perl, it probably has some type of basic problem imo. A common problem is a missing database index. Another is too much memory allocation, e.g. passing around a large scalar instead of a reference or overuse of objects (classical Java problem). It isn't always the case that you can fix the problem, but caching doesn't fix it either. At least understand the performance problem(s) thoroughly before adding the cache. Here's a fun example of a design flaw. It is a performance test sent to another list. The author happened to work for one of our competitors. :-) That may well be the problem. Building giant strings using .= can be incredibly slow; Perl has to reallocate and copy the string for each append operation. Performance would likely improve in most situations if an array were used as a buffer, instead. Push new strings onto the array instead of appending them to a string. #!/usr/bin/perl -w ### Append.bench ### use Benchmark; sub R () { 50 } sub Q () { 100 } @array = ( x R) x Q; sub Append { my $str = ; map { $str .= $_ } @array; } sub Push { my @temp; map { push @temp, $_ } @array; my $str = join , @temp; } timethese($ARGV[0], { append = \Append, push = \Push }); Such a simple piece of code, yet the conclusion is incorrect. The problem is in the use of map instead of foreach for the performance test iterations. The result of Append is an array of whose length is Q and whose elements grow from R to R * Q. Change the map to a foreach and you'll see that push/join is much slower than .=. Return a string reference from Append. It saves a copy. If this is the page, you'll see a significant improvement in performance. Interestingly, this couldn't be the problem, because the hypothesis is incorrect. The incorrect test just validated something that was faulty to begin with. This brings up you can't talk about it unless you can measure it. Use a profiler on the actual code. Add performance stats in your code. For example, we encapsulate all DBI accesses and accumulate the time spent in DBI on any request. We also track the time we spend processing the entire request. Adding a cache is piling more code onto a solution. It sometimes is like adding lots of salt to bad cooking. You do it when you have to, but you end up paying for it later. Sorry if my post seems pedantic or obvious. I haven't seen this type of stuff discussed much in this particular context. Besides I'm a contrarian. ;-) Rob
Re: When to cache
Perrin Harkins writes: To fix this, we moved to not generating anything until it was requested. We would fetch the data the first time it was asked for, and then cache it for future requests. (I think this corresponds to your option 2.) Of course then you have to decide on a cache consistency approach for keeping that data fresh. We used a simple TTL approach because it was fast and easy to implement (good enough). I'd be curious to know the cache hit stats. In this case, there was a high locality of access, so we got about a 99% hit rate. Obviously not every cache will be this successful. BTW, this case seems to be an example of immutable data, which is definitely worth caching if performance dictates. It wasn't immutable, but it was data that we could allow to be out of sync for a certain amount of time that was dictated by the business requirements. When you dig into it, most sites have a lot of data that can be out of sync for some period. I agree with latter clause, but take issue with the former. Typical sites get a few hits a second at peak times. If a site isn't returning typical pages in under a second using mod_perl, it probably has some type of basic problem imo. Some sites have complex requirements. eToys may have been an anomaly because of the amount of traffic, but the thing that forced us to cache was database performance. Tuning the perl stuff was not very hard, and it was all pretty fast to begin with. Tuning the database access hit a wall when our DBAs had gone over the queries, indexes had been adjusted, and some things were still slow. The nature of the site design (lots of related data on a single page) required many database calls and some of them were fairly heavy SQL. Some people would say to denormalize the database at that point, but that's really just another form of caching. Use a profiler on the actual code. Agreed. Add performance stats in your code. For example, we encapsulate all DBI accesses and accumulate the time spent in DBI on any request. No need to do that yourself. Just use DBIx::Profile to find the hairy queries. Adding a cache is piling more code onto a solution. It sometimes is like adding lots of salt to bad cooking. You do it when you have to, but you end up paying for it later. It may seem like the wrong direction to add code in order to make things go faster, but you have to consider the relative speeds: Perl code is really fast, databases are often slower than we want them to be. Ironically, I am quoted in Philip Greenspun's book on web publishing saying just what you are saying: that databases should be fast enough without middle-tier caching. Sadly, sometimes they just aren't. - Perrin
Re: When to cache
When you dig into it, most sites have a lot of data that can be out of sync for some period. Agreed. We run an accounting application which just happens to be delivered via the web. This definitely colors (distorts?) my view. heavy SQL. Some people would say to denormalize the database at that point, but that's really just another form of caching. Absolutely. Denormalization is the root of all evil. ;-) No need to do that yourself. Just use DBIx::Profile to find the hairy queries. History. Also, another good trick is to make sure your select statements are as similar as possible. It is often better to bundle a couple of similar queries into a single one. The query compiler caches queries. Ironically, I am quoted in Philip Greenspun's book on web publishing saying just what you are saying: that databases should be fast enough without middle-tier caching. Sadly, sometimes they just aren't. Every system design decision often has an equally valid converse. The art is knowing when to buy and when to sell. And Greenspun's book is a great resource btw. Rob
performance coding project? (was: Re: When to cache)
Rob Nagler wrote: Perrin Harkins writes: Here's a fun example of a design flaw. It is a performance test sent to another list. The author happened to work for one of our competitors. :-) That may well be the problem. Building giant strings using .= can be incredibly slow; Perl has to reallocate and copy the string for each append operation. Performance would likely improve in most situations if an array were used as a buffer, instead. Push new strings onto the array instead of appending them to a string. #!/usr/bin/perl -w ### Append.bench ### use Benchmark; sub R () { 50 } sub Q () { 100 } @array = ( x R) x Q; sub Append { my $str = ; map { $str .= $_ } @array; } sub Push { my @temp; map { push @temp, $_ } @array; my $str = join , @temp; } timethese($ARGV[0], { append = \Append, push = \Push }); Such a simple piece of code, yet the conclusion is incorrect. The problem is in the use of map instead of foreach for the performance test iterations. The result of Append is an array of whose length is Q and whose elements grow from R to R * Q. Change the map to a foreach and you'll see that push/join is much slower than .=. Return a string reference from Append. It saves a copy. If this is the page, you'll see a significant improvement in performance. Interestingly, this couldn't be the problem, because the hypothesis is incorrect. The incorrect test just validated something that was faulty to begin with. This brings up you can't talk about it unless you can measure it. Use a profiler on the actual code. Add performance stats in your code. For example, we encapsulate all DBI accesses and accumulate the time spent in DBI on any request. We also track the time we spend processing the entire request. While we are at this topic, I want to suggest a new project. I was planning to start working on it long time ago, but other things always took over. The perl.apache.org/guide/performance.html and a whole bunch of performance chaptes in the upcoming modperl book have a lot of benchmarks, comparing various coding techniques. Such as the example you've provided. The benchmarks are doing both pure Perl and mod_perl specific code (which requires running Apache, a perfect job for the new Apache::Test framework.) Now throw in the various techniques from 'Effective Perl' book and voila you have a great project to learn from. Also remember that on varous platforms and various Perl versions the benchmark results will differ, sometimes very significantly. I even have a name for the project: Speedy Code Habits :) The point is that I want to develop a coding style which tries hard to do early premature optimizations. Let me give you an example of what I mean. Tell me what's faster: if (ref $b eq 'ARRAY'){ $a = 1; } elsif (ref $b eq 'HASH'){ $a = 1; } or: my $ref = ref $b; if ($ref eq 'ARRAY'){ $a = 1; } elsif ($ref eq 'HASH'){ $a = 1; } Sure, the win can be very little, but it ads up as your code base's size grows. Give you a similar example: if ($a-lookup eq 'ARRAY'){ $a = 1; } elsif ($a-lookup eq 'HASH'){ $a = 1; } or my $lookup = $a-lookup; if ($lookup eq 'ARRAY'){ $a = 1; } elsif ($lookup eq 'HASH'){ $a = 1; } now throw in sub attributes and re-run the test again. add examples of map vs for. add examples of method lookup vs. procedures add examples of concat vs. list vs. other stuff from the guide. mod_perl specific examples from the guide/book ($r-args vs Apache::Request::param, etc) If you understand where I try to take you, help me to pull this project off and I think in a long run we can benefit a lot. This goes along with the Apache::Benchmark project I think (which is yet another thing I want to start...), probably could have these two ideas put together. _ Stas Bekman JAm_pH -- Just Another mod_perl Hacker http://stason.org/ mod_perl Guide http://perl.apache.org/guide mailto:[EMAIL PROTECTED] http://ticketmaster.com http://apacheweek.com http://singlesheaven.com http://perl.apache.org http://perlmonth.com/
When to cache
I'm interested to know what the opinions are of those on this list with regards to caching objects during database write operations. I've encountered different views and I'm not really sure what the best approach is. Take a typical caching scenario: Data/objects are locally stored upon loading from a database to improve performance for subsequent requests. But when those objects change, what's the best method for refreshing the cache? There are two possible approaches (maybe more?): 1) The old cache entry is overwritten with the new. 2) The old cache entry is expired, thus forcing a database hit (and subsequent cache load) on the next request. The first approach would tend to yield better performance. However there's no guarantee the data will ever be read. The cache could end up with a large amount of data that's never referenced. The second approach would probably allow for a smaller cache by ensuring that data is only cached on reads. In the end, this probably boils down to application requirements. RAM and disk storage is so cheap these days that the first method is probably fine for most purposes. However I'm sure there are situations where resources are limited and the second is more effective. What does everyone think? -- Milo Hyson CyberLife Labs, LLC