On Fri, 2008-06-13 at 08:35 +0200, Quintin Beukes wrote: > Hey, > > That's an idea (to first get the code optimized before squeezing it > into tomcat). I guess I just continued this way, as our current system > is using HttpURLConnection. So I wanted to get an idea of how it will > compare. > > I'll take your advice on running it in separate java apps? I assume > that is what you meant "direct connections", as in no proxied > connections via tomcat. >
Yep. > If someone had to take a wild guess, which of the examples of the site > can I use as a base if I'm targeting performance (a good balance > between throughput and high connection/second rates would be the > best)? > Most likely you want to run benchmarks against a site with a predictable routing overhead. If there are too many hops in between you can never be sure exactly what you are measuring: the performance of the HTTP stack or that of your Internet service provider's routing infrastructure. I guess 'localhost' is your best friend. Oleg > Q > > On Thu, Jun 12, 2008 at 11:02 PM, Oleg Kalnichevski <[EMAIL PROTECTED]> wrote: > > Quintin Beukes wrote: > >> > >> Hey, > >> > >> I made some basic servlets to test throughputs for HttpClient, > >> HttpCore and sun's sun.net.www.protocol.http.HttpURLConnection (or > >> something along those lines, I don't have an API at hand now). > >> > >> I basically just put up an apache session listening on port 8580, and > >> then Tomcat on port 80. In tomcat I dumped 3 servlets, each using one > >> of the 3 methods to retrieve a 51kb page from the apache session. > >> > >> The benchmarks was done with "ab" with 10 second timeouts. The results > >> are as follows (in byte readings, not bit readings). > >> > >> Doing a direct request to apache I get a throughput of about 33mb/s. > >> Using HttpURLConnection, 4.7MB/s. > >> Using HttpCore 3.6MB/s > >> Using HttpClient 6.03mb/s > >> > >> The request times were in general best with apache/HttpClient, maxing > >> at 38ms for the longest request. > >> > >> This was done at a concurrency of 1 connection. Pushing this up to 5, > >> the HttpCore's connection times stayed the same, but throughput > >> dropped to 980kb/s. HttpClient's connection times now came to the same > >> as HttpCore (230ms for longest request), but it's throughput increased > >> to 7.1mb/s. > >> > >> This is all good and well, but my question is this. Are there any ways > >> I can increase these throughputs? Before just jumping into > >> optimizations suggestions and referals to documentation, could someone > >> perhaps just enlighten me into why these speeds are so different. And > >> especially so low? > >> > > > > I do not want to jump into conclusions here, but I seriously suspect > > something is wrong with your test cases. In particular I do not understand > > why you are using Tomcat to host the code your are testing. What is the > > point of that? > > > > I made an extensive comparison of HttpCore vs HttpClient (both versions 3.1 > > and 4.0). HttpCore can easily achieve a throughput of 10,000 (small) > > requests per second on a fairly average hardware. > > > > http://wiki.apache.org/HttpComponents/HttpClient3vsHttpClient4vsHttpCore > > > > > >> I am obviously using HttpCore wrong for it to be so slow (I mean, > >> HttpClient is built on core, so it's impossible to be faster). It's > >> just that the documentation for these are sparse, and I'm having a > >> very hard time finding my way around. > >> > >> And beyond this, Tomcat itself gives me fantastic speeds reading > >> static data. Where are the bottlenecks? > >> > >> The HttpCore setup I used was the "Basic GET Requests" example > >> available on the site. I modified it a bit to allow URL input and > >> target server specified through parameters passed to the servlet. > >> Beyond that I didn't change anything. My assumption is that > >> connections aren't being reused. In fact, I know this is the case, so > >> it logically and obviously removes the validity of the above > >> benchmarks for HttpCore. I'm not going to remove them though. > >> > >> I would really like to get throughputs similar to direct Apache > >> connections (at least 33+ mbyte/s, which is round about the general > >> limit of hard drive transfer rates). Full ethernet throughput is > >> preferable and I'm sure possible. Downloading large files I get speeds > >> of up to 54mbyte/s with Apache. I'm sure this is limited only to the > >> hard drive I tested with, as it's a gigabit connection, and this is > >> just less than half it's full potential. I haven't done tests using > >> memory transfer. Will try it when I can diagnose the bottlenecks I > >> currently have, and then use them as measurements to gain insights > >> into the speed of the program itself by taking as much of the other > >> bottlenecks in the system away as possible (in this case the > >> limitation of data transfer rates from the hard drives). > >> > >> The three benchmark servlets I used are here: > >> HttpClient: http://quintin.dev.junkmail.co.za/httpclient/Bench.java > >> HttpCore: http://quintin.dev.junkmail.co.za/httpclient/BenchCore.java > >> HttpURLConn: http://quintin.dev.junkmail.co.za/httpclient/BenchURL.java > >> > >> And a utility class I use in the above: > >> http://quintin.dev.junkmail.co.za/httpclient/StreamBinder.java > >> > >> Any insights would be greatly appreciated. > >> > > > > I'll try to look at the benchmark code over the weekend. But I would > > _strongly_ suggest measuring the performance of each respective framework by > > using direct connections to the target server. Once you have established the > > baseline you can work on optimizing the code when running inside a servlet > > container. > > > > Oleg > > > >> regards, > > > > > > --------------------------------------------------------------------- > > To unsubscribe, e-mail: [EMAIL PROTECTED] > > For additional commands, e-mail: [EMAIL PROTECTED] > > > > > > > --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
