4 ms·
Rough Estimates of the Dollar Cost (of Memory) of Scaling Web Platforms - Part I
- sadiq 17y agoJetty's initial heap size can be tuned. Setting the initial heap to something small (say 16mb or 32mb), while having a decent max heap, would change those graphs a fair bit.
- davidw 17y agoCool, thanks. I'll collect some of this advice and then update the article. Part of what made doing this difficult is that some things I know reasonably well, and others not so well, so I had to trust some of the out of box settings.
- profquail 17y agoI would've liked to have seen Apache/Mono (mod_mono) in there as well. I've been trying to make sure my ASP.NET sites are cross-platform just in case I want to move them over to a Linux server sometime, but I haven't had a chance to do any load testing on them yet to see how it compares to Windows Server or any of the other Linux app platforms.
- davidw 17y agoIf you can get me an easy-to-follow setup/config, I'd be happy to include it.
- aaronblohowiak 17y agoApples:Oranges. Why not compare Rack or Sinatra to PHP, which would be a much closer comparison. Or, perhaps try CakePHP or one of the 1000 other php frameworks out there. Similarly, should be including web.py instead of Django. Comparing Rails to Jetty is pointless! Compare Rails to Spring/Hibernate.
- skorgu 17y agoThere are a couple of different fruits being compared (and maybe making this more plain would avoid this kind of critique): Apache, Nginx, the static servers are baseline connection costs. PHP, Jetty/JSP, Mochiweb templated are one tick up, baseline connection plus the template cost. I'd like to see Tornado included here personally. Rails and Django are full stacks. It would be nice to see Cake or Symfony included here and I personally would like to see Compojure as well but that's much more niche than the others. Edit: If nobody beats me to it I'll do the date/time template in Tornado and Symfony this evening.
- davidw 17y ago> Edit: If nobody beats me to it I'll do the date/time template in Tornado and Symfony this evening. Thanks, I'm looking forward to seeing it!
- skorgu 17y agoEdit: Rather than fighting with escaping code in HN, here are scripts to setup symfony and tornado examples: http://gist.github.com/275785 http://gist.github.com/275785 The symfony one requires some more loving to the apache setup which is well documented (http://www.symfony-project.org/getting-started/1_4/en/05-Web-Server-Configuration http://www.symfony-project.org/getting-started/1_4/en/05-Web...). The script suggests the frontend_dev.php to make sure things are set up right but for actual performance benchmarking you'll have to tweak your apache.
- davidw 17y agoOk, I'll have a look. By the way, you have no contact information or anything in your profile, so I have no way of getting in touch with you.
- megamark16 17y agoI'd like to see how a fairly straight forward webapp running on different web app frameworks compare. Lets say you write a simple polling app (similar to the Django tutorial example) using Django, Rails, ASP.NET/Mono, and Node.js :-) Write the app to use whatever features you can within the framework but with matching output, perhaps sans caching, and run them all on similar architecture where possible (apache/passenger maybe?) then see how those load tests compare. If nobody else does this than I might, I'd be interested to know what that looks like.
- skorgu 17y agoYeah, I think this would be a great idea. I'd like to see a set of traffic generators that exercise actual functionality not just ab/pageload tests too, that works more of the stack. It should be reasonably easy to automate a deploy/test run in EC2 too, then you could do a big batch of comparisons with configuration changes, different db backends, tuning parameters, etc. Sort of like the programming language shootout but for webapps.
- davidw 17y agoFor everyone saying: "I'd like to see ...", give me some instructions on how to set it up on Ubuntu, and I'll see about adding them. That is part of the reason why it's a "Part I".
- houseabsolute 17y agoI'd like to see non-3D graphs. ;)
- davidw 17y agoOops, fixed the one that was.
- pquerna 17y agoThis is just wrong. The Virtual Memory Size (VSZ as he calls it in the article) is not what is important. Measuring actual memory preasure and costs on modern Linux is pretty non-trivial, just look at this wiki page about Spam Assassin: http://wiki.apache.org/spamassassin/TopSharedMemoryBug http://wiki.apache.org/spamassassin/TopSharedMemoryBug He needs to measure the real size of a multi-process daemon like apache by using a script like the ps_mem.py: http://www.pixelbeat.org/scripts/ps_mem.py http://www.pixelbeat.org/scripts/ps_mem.py That can measure the real memory use of each daemon, and while I don't think the overall picture will change much, measuring VSZ is just inaccurate.
- davidw 17y ago> while I don't think the overall picture will change much, measuring VSZ is just inaccurate. I agree... which means that your "it's just wrong" comment is inaccurate too, because the general relationship of the various systems to one another, in terms of memory usage, is not likely to change much: it's obvious that one way or another, Rails doing complex dynamic stuff with templates is going to use more memory than nginx serving static files, with other things falling in between. Thank you, in any case, for the pointer to ps_mem.py. I am using a similar script, but it doesn't look at smaps, so I'll improve the one I have or use theirs.