W3C Jigsaw


WARN ! This page has not be updated since the 1.0alpha3 release, we do plan to do so real soon, stay tuned. To get some ideas as to how Jigsaw compares to a C based HTTP server, you may want to check the HTTP Performance Paper (which, even though not exactly on the same subject, provides a fair idea of relative performances).


Performance Evaluation

This document describes the current performance achieved by Jigsaw, and compare them to other servers. This paper is not intended to prove that Jigsaw is the fastest server (which it is not), it just tries to point out that writing a server in Java is a reasonable thing to do.

The second section of this document will describe the current bottlenecks, and hilight some of the possible work arounds (which will hopefully be included in the next Jigsaw release).

Numbers

Caveat

Benchmarking HTTPD servers is definitely a difficult task. They are at least two difficulties here:

The benchmark will usually depend on things such as your hardware and your system, but also, the load of the machine you are using for the benches, the kind of filesystem the server runs in (is it NFS, AFS, does it uses some automounter ?), and a number of other parameters that are difficult to measure. The numbers we will give in this section have all been obtained under the same global configuration. This is a SunSparcstation 5, 64Mb of RAM, running solaris2.5, all files are served through a read/write AFS volume. I am probably lacking a number of interesting parameters in the latter description, that's why I will focus on relative numbers (i.e. how Jigsaw performs relatively to other servers), rather then absolute numbers. You shouldn't even try to scale these numbers in any ways: some servers speedup won't be linear with the power of the machine it runs on (e.g. if server foo is 2 times faster then server bar on a sparc station 5, the relative performance of foo and bar on a sparc 20 may not be the same).

How are these numbers collected ? The number we will present are by no mean related to what a typical server will experience in some real life conditions. For example, in our benchmarks, the client and the server programs are always run on the same machine. This probably affect the request rate, however, it still gives us a good indication of the CPU consumption of the server, which is what we are really interested in. To collect these numbers, we have used the ptester program which is a simple C hitter provided with the phttpd server.

The benchmarks

For each of the servers we measure, we run the following four benchmarks:

Single document retreival
This first bench gets the same document in sequence, and prints out the number of requests achieved for a given duration of time. For those of you familiar with ptester, the command line is as follows:
ptester -t60 <document>
We run this bench in two variants: first with a 4kb file(1.a) and 40kb file (1.b)
Multiple document retreival
The second bench puts a little more load on the server, by requesting three documents in parallel. The three documents are respectively of 4kb, 10kb and 40kb. For those of you familiar with ptester, the command line is as follows:
ptester -t60 -rN <doc1> <doc2> <doc3>
We run this benchmark in two variants: first with N=2 (2.a) 2 simultaneous hitters for each document, for a total of 6 simultaneous request flows and with N=10 (2.b) for a total of 30 simultaneous connections.
Keep-alive enabled bench
This last bench will only benefit to those servers that implement the HTTP keep-alive feature. Each connection will be used to run multiple requests. For those of you familiar with ptester, the command line is as follows:
ptester -t60 -kN <document>
The bench is only run for a single document of 4k, and for a k value of 5, it is test 3.
Long live bench
This is to measure (or get some idea at least) the impact  of the Java garbage collector on Jigsaw. It's a replay of test 2 for a duration of ten minutes, which give us our final benches 4.a and 4.b

Results

While running the benchmark, Jigsaw is configured to generate the appropriate (CommonLog) log files; no ident protocol is run.

We have obtained the following results:

Benchmark results in requests per second
version 1.a 1.b 2.a 2.b 3 4.a 4.b
CERN-server 3.0 11 10 11 11 NA 11 10
Apache 1.0.2 35 24 30 30 NA 33 30
Apache 1.1.1 42 30 35 33 85 34 31
phttpd 0.99.70.1 46 37 39 38 4 39 37
Jigsaw 1.0alpha1 22 18 19 9 26 19(*) 8(*)
Jigsaw 1.0alpha3 22 18 19 18 26 19(*) 18(*)

Due to some problems with the Java garbage collector, benches marked with a * have been run with 8 Mb of memory allocated to the Java interpreter rather than the default 1Mb, we'll come back to this.

Cells marked NA (not available) don't make sense for the given server (they don't support kepp-alive).

Lessons

A number of lessons can be deduced from the above performance numbers. Some of them are related to the current Jigsaw implementation, others are related to the current Java implementation.

All these lessons should be taken as hypothesis, only some experiments will tell the truth abouth these numbers.

Jigsaw implementation problems

Don't forget that Jigsaw current implementation is flaged as being alpha. This means a number of things, among which the fact that Jigsaw code has not yet been profiled and optimized. Jigsaw design should provide the basis of a fearly efficient httpd implementation. In particular, features such as authentication, logging, etc. which are not measured by the above benches should prove fearly efficient (the main reason for this assertion lies in the clever caching scheme Jigsaw uses for resource meta-informations).

However, what are the above numbers telling us ? First of all, benches 1.a and 1.b shows that Jigsaw throughput (the number of bytes/second it can emit) is not that bad. We did some experiments with a native shuffler to try to infer the cost of the actual file transmissions across the wire. This would run as follow: the Java part of the server would compute a reply, but would delegate the actual sending of the bytes to an efficient C multi-threaded process by using a Java native interface to the UNIX sendmsg system call. Although this gave us much better numbers on solaris 2.3, the results on solaris 2.5 were quite disapointing (the speed-up would totally disapear). The good news, however, is that the Java threaded IO performances are not that bad, and are probably not at this time the bottleneck. We do still, however, expect a significant boost in these number when the Java runtime will be able to use solaris native threads (although, at this stage, this should be taken as an intuition only).

As of 1.0alpha1

The real lesson from these numbers comes from the benches 2.a and 2.b. Jigsaw performance seems to be very sensitive to the number of simultaneous connections. After some quick investigation, our guess is that the main problem with the implementation is that it contains too many central resources, and that mutliple threads are mostly blocked trying to acquire them. At this point, we have detetected at least three of these:

The client pool
This object is the central part of the server: it dispatches incomming connections to client objects which will then handle them. The current implementation of the client pool has some nice features (for example, it keeps a pool of threads, ready to run on behalf of any new incomming connections). However it's current implementation locks the whole object during the selection process of the client (which is poorly implemented right now). This means that even if the server threads accept(2) connections as fast as they arrive, it will block to acquire the lock on the client pool before being able to dispatch the connection.
The event manager
Before starting to read any incomming requests, client objects set a timer in order to avoid maintaining idle connections for too long. Once the request has been read and parsed, they then arm a second timer to control the duration of the request processing. All these timers are set through a single event manager, which is locked during both timers setting and timers cancelation. To be more practical, if 10 simultaneous connections are made to the server, then ten client objects will try to lock the event manager at the same time, competing to acquire the appropriate lock.
The logger
The logger is the central object in the server that logs all requests. As for the above resources, deficiency in its current implementation (this logger, for example, doesn't bufferize output) makes it a bottleneck.

Note that generally these bottlenecks exist only because the underlying implementation of the objects are not yet optimized (at this time, I don't think of them as being a design problem). To get an idea of how well Jigsaw would perform if these bottleneck were eliminated, I have put up together a version of Jigsaw that :

For the client pool, nothing simple can be done at this time. However, just with the two other optimizations, we already get  28 requests/second for bench 2.a and 22 requests/second for test 2.b. The really interesting thing is to notice how the gap between these two numbers disappear (which means we really had cotention problems).

This is for the Jigsaw implementation. We hope that by the next release all these contention problems will be solved (the three above are on the top of the TODO list). However, as we mentioned in the preamble of this section, there is another set of problems related to the current Java implementation.

Jigsaw 1.0alpha3

Most of the 1.0alpha1 bottlenecks have been reduced, here is a brief overview of what has happened to the three previous ones:

The client pool
Has been entirely rewritten. The new client pool tries to be pretty smart by running in one of four modes:
The event manager
The event manager jobs has been reduced to keep track of requests that take to long to request. Getting rid of the keep alive timeout has basically reduced by half the load on the event manager.
The logger
The logger now uses a n optional buffer, that speeds up the critical section were it emits log entries.

It should also be noted that Jigsaw HTTP API has been totally redesign, and is much more efficient the the one included in 1.0alpha1.

The result of these changes is mainly seen through the much better handling of simultaneous connections that Jigsaw now provides.

Java implementation problems

Although Java comes from an old (well, at least three years old) project, its current implementation is still pretty new. They are a number of problems with this implementation that need serious work to be fixed.

First of all, let's start with the byte-code interpretation, since this is what people are the most scared with. As our numbers show, Jigsaw out-performs the CERN server by far, on nearly all benches (except for benches 4.a and 4.b, which we will come to later). The CERN server is written in C. So at least, this proves that the interpretation overhead of Java can be dealt with by a better design. I would guess that even when JIT compilers come out that are able to speed up Java code by 10 times, compiling Jigsaw with them won't make it ten times faster, just because byte-code interpretation may not be the overhead in Jigsaw.

The most serious problems with the current JDK is that it doesn't come with the appropriate tools for profiling and detecting the hot-sopt in the server's code. The three bottlenecks that we have exhibited in the previous section were found by simply printing the thread statuses at various point in time under heavy load. Getting the right tools should help a lot optimizing Jigsaw implementation.

Another problem with the current Java implementation is the thread-safe libraries. If you use the StringBuffer class, for example, you will slow down your application because most of its methods are synchronized, while at some point, you know your thread is the only one dealing with some instances of this class. The StringBuffer class is the first concern here, since you usually want to write code such as:

String x = ... ;
String y = ... ;
String xy = x + y ;

What really happens in the third line here, is that you implicitly allocate a StringBuffer instance, initially filled up with the content of string x, and then you append to it the content of string y. Then you get the StringBuffer as a String to get back the result that you assign to xy. Note only have you allocated a StringBuffer object, but you had to allocate a monitor for it (locking the global monitor cache for some time), and then call a synchronized method in order to do the concatenation. Depending on the conditions in which the above code is running you can work-around it by implementing your own concatenation operations in your object (this is what Jigsaw MIME parser does, for example).

Now, we still haven't exlpain the purpose and the result of benches 4.a and 4.b. These are meant to see how the Java garbage collector react when its hosting application is under heavy load. [WARNING: following explanations are not bullet-proof, also they have good roots.]. The general garbage collection process of Java can be divided in a two step process:

  1. The garbage collector runs, and detect all the objects that are currently garbage (i.e not reachable from any of the global roots). These objects are marked as such, but not yet collected.
  2. The finalizer runs through all the previously marked objects and finalize them by invoking their finalize method.

Until an object has been finalized, its memory cannot be reused. Both stages one and two can run either asynchronously (i.e. when the system is idle) or synchronously (i.e. because they are required to run in order for the process to continue its job). The problem we are experiencing in benches 4.a and 4.b is the following: as the server is never idle during the bench, the asynchronous version of the two stages is never able to run, so  the server relies entirely on the synchronous version of them. What really happens here is best depicted by running the server with the verbosegc option turned on:

www24:Jigsaw$ java -verbosegc w3c.jigsaw.http.httpd 
loading properties from: /afs/w3.org/usr/abaird/Jigsaw/config/httpd.props
*** Warning : no logger specified, not logging.
[httpd]: listening at:http://www24.w3.org:9999
<GC(async): freed 972 objects, 29712 bytes in 34 msec, 95% free>
<GC(async): freed 2 objects, 16 bytes in 29 msec, 95% free>
<GC: freed 3877 objects, 288960 bytes in 26 msec, 34% free>
<GC: freed 736 objects, 285608 bytes in 23 msec, 54% free>
<GC: freed 933 objects, 76040 bytes in 65 msec, 38% free>
<GC: freed 1926 objects, 150320 bytes in 27 msec, 24% free>
<GC: synchronously running 408 finalizers>
<GC: freed 0 objects, 0 bytes in 64 msec, 24% free>
<GC: expanded object space by 8192 to 847048 bytes, 25% free>
...[more traces omited]

What happens here is that the asynchornous GC never gets a chance to run (as expected). The synchronous GC tells us about its good work, however free memory ratio never improves: the finalization process doesn't run. As for the GC we don't expect the asynchronous version of it to run, but the synchronous finalizer falsly indicate that it has finazlied some objects (while in fact, the synchronous finalizer was probably unable to finalize them, because the asynchronous finalizer was lying around).

For real fun, here are two other traces collected under exactly the same circumstances:

www24:Jigsaw$ java -verbosegc w3c.jigsaw.http.httpd 
loading properties from: /afs/w3.org/usr/abaird/Jigsaw/config/httpd.props
*** Warning : no logger specified, not logging.
[httpd]: listening at:http://www24.w3.org:9999
<GC(async): freed 972 objects, 29712 bytes in 34 msec, 95% free>
<GC(async): freed 2 objects, 16 bytes in 27 msec, 95% free>
<GC(async): freed 1424 objects, 92224 bytes in 61 msec, 86% free>
<GC(async): freed 179 objects, 2368 bytes in 35 msec, 86% free>
<GC(async): freed 520 objects, 41088 bytes in 33 msec, 86% free>
<GC(async): freed 72 objects, 960 bytes in 38 msec, 86% free>
<GC(async): freed 325 objects, 25680 bytes in 40 msec, 86% free>
<GC(async): freed 45 objects, 600 bytes in 31 msec, 86% free>
<GC(async): freed 1430 objects, 112992 bytes in 37 msec, 86% free>
<GC(async): freed 198 objects, 2640 bytes in 31 msec, 86% free>
<GC(async): freed 1950 objects, 154080 bytes in 37 msec, 86% free>
<GC(async): freed 270 objects, 3600 bytes in 32 msec, 86% free>

This run is pretty ideal. For some reasons, the asynchronous GC manages to run, and the memory is taken care of nicely (Jigsaw achieved 30 requests/second during this test).

The last configuration I have encountered is the following:

www24:Jigsaw$ java -verbosegc w3c.jigsaw.http.httpd 
loading properties from: /afs/w3.org/usr/abaird/Jigsaw/config/httpd.props
*** Warning : no logger specified, not logging.
[httpd]: listening at:http://www24.w3.org:9999
<GC(async): freed 972 objects, 29712 bytes in 35 msec, 95% free>
<GC(async): freed 2 objects, 16 bytes in 27 msec, 95% free>
<GC: freed 9208 objects, 707008 bytes in 33 msec, 84% free>
<GC: freed 7233 objects, 571648 bytes in 34 msec, 82% free>
<GC: freed 5834 objects, 460704 bytes in 34 msec, 81% free>
<GC: freed 4550 objects, 359520 bytes in 34 msec, 80% free>
<GC: freed 4485 objects, 354384 bytes in 36 msec, 79% free>
<GC: freed 4420 objects, 349248 bytes in 41 msec, 78% free>
<GC: freed 4355 objects, 344112 bytes in 40 msec, 77% free>
<GC: freed 4355 objects, 344112 bytes in 42 msec, 76% free>
<GC: freed 4355 objects, 344112 bytes in 44 msec, 75% free>
<GC: freed 4290 objects, 338976 bytes in 45 msec, 74% free>
<GC: freed 4160 objects, 328704 bytes in 46 msec, 73% free>
<GC: freed 7222 objects, 309792 bytes in 41 msec, 79% free>
<GC: freed 6319 objects, 297248 bytes in 31 msec, 83% free>
<GC: freed 8631 objects, 647424 bytes in 32 msec, 84% free>
<GC(async): freed 2557 objects, 129320 bytes in 42 msec, 86% free>
<GC(async): freed 212 objects, 2848 bytes in 32 msec, 86% free>

The memory keeps getting down, slowly. If the test continues for a while, the interpreter will fall back in the first situation presented above (i.e. trashing). However, whenever the test is stopped (as in the last two lines above) the asynchronous GC normally runs and get back all the memory.

Conclusion

What can you expect - with regard to performances - in the next Jigsaw release ? As said above, the Jigsaw implementation should improve, and provide better performances.

However, some real speed-up can be expected by better Java implementations in the next six monthes. By this time, the GC will hopefully be debugged, and on solaris at least, we can expect the native solaris threads to replace the Green thread package that is currently being used.

Send suggestions, comments, criticisms, corrections to Jigsaw Team



Jigsaw Team.
$Id: performance.html,v 1.14 2002/08/07 12:16:50 cbournez Exp $