Slashdot Mirror
Should Servers be Mono-Process or Multithreaded?
An anonymous reader wonders: "How would you design the fastest possible Linux-based server application today? A few years ago, the thinking was that multi-threading was not the way to go — instead, high-performance servers used an event-driven, mono-process model (consider lighttpd and haproxy). However, things have changed. Today CPUs have dual cores, and over the next few years this is only likely to increase. Also, the 2.6 Linux kernel has made multi-threading much more efficient. So I'm wondering, does Slashdot think that modern high performance server software should be designed to be multi-threaded, or does it still make more sense to use an event driven, mono-process architecture, despite the advances in the Linux 2.6 threading and the arrival of multi-core CPUs?"
Check out the C10K page for a very detailed discussion about this.
Did you ever notice that *nix doesn't even cover Linux?
Well first, you probably should keep things simple and just buy nice hardware. Most servers sit idle most of the time anyway. If you truly do need the perfornamce though...
Have 1 process per node. I mean "node" in a NUMA sense. 64-bit AMD systems have one node per chip package. Other PCs (except exotic stuff) have one node for the whole system. Lock your processes to separate nodes, so that they all get local memory. If you don't do this, at least remember to use the new system call for moving pages from one node to the other. (eh, "move_pages" if I remember right -- see unistd.h in the kernel source)
You'll need extra threads to do disk IO. Not counting those: On each node, have at least 1 thread per bottom-level (usually L2 or L3) cache, but not more than 1 thread per virtual core (hyperthreading thing). If you go with 1 thread per physical core but have virtual cores (hyperthreading) enabled, lock your threads to virtual cores that don't share phycical cores.
A lot of this should be configurable. Hopefully you'll make an easy way to automatically determine the best configuration, writing out the appropriate config file so that manual config hacking is not required to get the best performance.
You forgot multiprocess. Like anything in software the answer is, "it depends on the application". But one of the most overlooked and frequently very important factors that affects performance is cache locality. If the CPU has to fetch something from main memory (or heaven forbid it actually has to drudge it up from disk) the program has to wait. That wait time is often much much greater than the execution time of the target code. Aside from simply writing small code (that only get's you so far), one way to get better cache locality is to break up your processing into a pipeline. Mail servers frequently do this. One process will accept connections do some sanity checking and write the message to another process. The next process juggles addresses for routing and writes it to another process. That process might then work on delivery either locally or remotely. What happends (or what is supposed to happen under high load) is that one process becomes hot and processes as many messages as it can until the buffer to the next process is full. Then the next process runs processing all of those messages until it either runs out of stuff to process or cant write anything more to the next process in the pipeline. If you have multiple cores / CPUs this scales pretty well too.
But again, "it depends on the application". The above pipelining method only performs well if you're processing items in an assembly line fashion. If you're an HTTP proxy server you wouldn't want that model. You would probably want a single process libevent type of thing. I have some code that doesn't use either of those models. It's a multiprocess model but event driven with *everything* in shared memory. It's very close to a multithreaded model but I needed security context switching. Also, contrary to popular belief threaded servers are slower than equivalent multiprocess servers. So in-general, the benifit of a multithreaded server is pretty much just about convenience for the programmer. Since you can acheive the same effect by just creating an allocator from a big chunk of shared memory mapped before anything is forked, there's very little reason to use threads at all.
Since you didn't say what kind of server you're building, I'm going to assume:
- that you're building a custom-purpose, client-server or message processing application,
- it needs to be highly parallel to be efficient
- the language is C, C#, or C++, and not Java (process-based servers in Java?)
I have done this before, using both processes and threads, for the same application. Consider the impact of application faults on your design, and then consider how hard it will be to create thread-safe code.
o A highly multithreaded server, where threads are performing complex and/or memory demanding tasks, will be susceptable to complete failure of all running jobs on all threads, if just a single thread SEGFAULTs. And despite your best testing efforts, complex code (1M+ lines) will at some point, somewhere, fail.
o Threaded code must be thread safe. Static variables, shared data structures, and factories all previously accessed through a singleton must now be protected with guard functions and semaphores. Race conditions need to be considered. Design for this up front. It will be much harder to add it later.
The Project
I worked on a team which added an asychronous processing engine to a web application. The engine was responsible for performing memory and time-intensive financial analysis and reporting for 16,000 accountants, so that they could close a large company's financial books. Unlike a webserver triggered by on-line end users, this engine is triggered by events in the company's financial database: once the database raises the "ready" flag, this engine begins running as many reports as it can, as fast as it can, on behalf of the 16K users. The analysis and report code was 2 million lines of C++, running on AIX.
Process implementation
The initial implementation used processes. A dispatcher job monitored the database for the ready flag, and then forked children of itself to analyze slices of the data, and generate the reports. One child job was used for each analysis and report pair, and the manager controlled how many jobs ran in parallel, maintaining a scoreboard of which jobs succeeded, and which failed.
Due to the complexity of the system, failures (core) occasionally occurred. The monitor would record this, retry the failed analysis up to 3 times, and keep a uniquely named core file of the event. Other analysis reports would continue to be generated, otherwise unharmed by the thrashing thread. Approximately once every 90 days, the development team would collect the few cores generated, use the gbd/xldb debugger to determine the cause of failure, and correct the fault.
The downsides of this? The solution was slowed because couldn't re-use resources like database connections (they were destroyed with each process), and more memory was used than need be. DB2 caching helped somewhat, but potential performance improvements remained.
Threaded implementation
In a large company, there are IT standards, and one of the standards at my company is that applications shall never, ever, ever fork(), even if running on a large dedicated machine. After losing the fight against this, my team re-architected the report engine. Largely
the same as the previous, the new engine waits for the "ready" signal, and then spawns pthreads (POSIX threads) as workers to analyze the data and generate the report. In theory, it was robust.
The alpha version of this solution immediately failed (cored) during testing. We neglected to identify the less obvious non-thread-safe code in the application, and failed to identify several race conditions. Unlike previous failures, this faults were total: a SEGFAULT in code on one of 20 threads would halt the entire application. And the corefile generated was now huge - it contained a snapshot of memory for all 20 running jobs, instead of just the one of interest.
Extensive root-cause analysis, design, and restart management solved this, and the current version is as robust, and a good bit faster, than the previous. At a significant price.
Beware: I believe all are created equal, and have the right to life, liberty, and the pursuit of happiness.
I cannot speculate, but I can look at what people are doing today. One thing that I have noticed, is the widespread research into, with compelling arguments, for massively multithreaded programming techniques. See Erlang for example. It is designed right from the beginning for this sort of problem - high throughput, high reliability, high uptime telephony networks.
As a rough benchmark, someone's got this.
That's an order of magnitude increase in "performance" (depends on what you mean by performance". I thought I'll do a casual informal test of my own, with a decent static file size (instead of the 1 byte used in that benchmark)
Server Software: Yaws/1.56
Document Length: 402 bytes
Concurrency Level: 500
Time taken for tests: 8.480740 seconds
Complete requests: 5000
Requests per second: 589.57 [#/sec] (mean)
Time per request: 848.074 [ms] (mean)
Server Software: Apache/2.0.54
Document Length: 402 bytes
Concurrency Level: 500
Time taken for tests: 29.787216 seconds
Complete requests: 5000
Requests per second: 167.86 [#/sec] (mean)
Time per request: 2978.722 [ms] (mean)
Output edited to get past lameness filter.
Err crap, I could have sworn the first time I tried this, when Yaws was first installed, its performance was worse! Oh well, perhaps it's something I've inadvertently done since then. Could have been due to my computer reboot (this is a desktop PC). It seems I've proven my point, although I was trying to disprove it. Standard caveats regarding benchmarks apply. Both servers are default Ubuntu installs with no configuration changes - I didn't compile anything manually.
Additionally it has also been noted that:
Well, that's where it could be headed anyway - a multiprocessor system with green threads (ie simulated threads, like Java ones) implementing massive concurency and redundancy. Some prototypes for systems like this are already available, and being used. Cheers.
start-rant:
/rant.
Threads are useful, that's granted - but it would seem a lot of people are trying to convert wholesale over to this threading model just for the hell of it, running along with the apparent reasoning that threading is "lighter" than processes. Maybe threads are lighter/cheaper on Windows systems - but a Unix system with copy-on-demand paging forking/process system is _DESIGNED_ to handle processes. Right now a lot of the time threads are a hack. Unix and processes work nicely together.
As for "maximising" available resources, well don't forget there's typically another couple of dozen processes running on any give Unix setup, more so on a multi-user multi-purpose machine (let's say WWW, email and DNS setup - throw in SpamAssassin for lots of fun) there's no shortage of available processes to use up a CPU. On a monolithic system where it's running only one process, sure, threads become useful there to spread the load.
My gripe basically boils down to a lot of people going along and choosing to use threads rather than forking because they think that it's "cool" or (supposedly) "lighter" - not because they've done any real world testing/checking. Remember, Unix was built around the idea of many small processes/programs working together, so that'd tend to naturally allow usage of multiple CPUs without any exotic hacks.
however, it's easier conceptually to write a threaded server, it's more natural to write, and you just launch a single thread per connection. unfortunately, currently, this doesn't scale (see Why Events Are A Bad Idea (for High-concurrency Servers) http://www.usenix.org/events/hotos03/tech/vonbehre n.html for an argument that thread implementations, and not their design, are the issue).
the former method can handle thousands of simultaneous connections with high throughput, even on a decent workstation; the latter cannot. threads simply have an inherent overhead that cannot be eliminated.
i've actually been working on writing a non-portable insanely fast httpd in my spare time (svn co svn://parseerror.dyndns.org/web/) over the past few weeks as a way to explore non-blocking I/O + epoll() and it performs very well (~600% faster conns/sec than a traditional fork()ing server (which i wrote first)).
for further discussion see The C10K Problem http://www.kegel.com/c10k.html which goes in-depth on these very subjects
What should I do?
This debate hits home with me. I wrote a server daemon to handle the SMTP and POP protocols, and when I first started out I had to make a choice. The choice I made back then was to use a threaded model. The way it works is I spawn X threads which collectively use blocking calls to the accept() function. Each thread will only return from accept() once they have been assigned a new connection by the kernel. For performance I spawn the threads ahead of time. This architecture was a mistake. The issue is that I have to spawn a seperate pool of threads to listen on port 25 (SMTP), port 110 (POP), port 465 (SMTP over SSL) and port 995 (POP over SSL). With this model if I could end up with extra threads listening on port 25, when I need more threads listening and processing connections on port 465. This problems leads me to overcompensate by spawning _extra_ threads just in case. Of course this strategy wastes resources as now extra threads eat memory without benefit.
To address the SEGFAULT issue, ie one rouge thread taking the whole system down, I also fork multiple processes. In my case I fork 12 processes with 128 threads each. If one process gets killed by a SEGFAULT, the remaining processes continue to work. When I first launched the system, and it faced a torrent of email... 100K+ messages a day, I would have about one process die every 24 hours. With careful debugging work, I've gotten the code stable enough now that I haven't lost a process in about 9 months.
My theory when I first wrote this code was to leave scheduling to the kernel. I figured that if a thread was blocked waiting for IO data the kernel wouldn't schedule time slices for it. This meant those extra threads sat in the background waiting, but not using CPU time. I am starting to wonder whether this is a good theory? I am considering switching to a different model (more on that in a second), but am not sure which one is best? By the way, the reason each process has so many threads is for DB connection pooling. Each process gets 8 DB connections which are shared between the 128 threads. Each process also has its own copy of the antivirus database. I know its possible, but trying to share DB connections and data between processes is much more difficult.
I plan to refactor this code soon, and have been struggling with what to do. I am curious to hear the thoughts of others?
The current plan is to move to a model where I spawn a single thread for each port. When these listening threads have a new connection, they dump the socket handle, and the the protocol into a buffer. I would then also spawn a pool of worker threads which read the incoming connections out of the buffer. Using semaphores and reflection these worker threads would pickup incoming connections and feed them to the right function depending on the protocol. I think this model would work much better than what I have now, but is this the best option?
The other option is to create system where I spawn only 8 worker threads (or some similar number). This pool of 8 threads then uses epoll() to find out which sockets need attention and address them accordingly. The problem with this model is that if an incomplete message is receieved, the thread couldn't process all the way into the output stage. Instead the data would need to be stored until the message sending was complete. Let me give an example, the thread might get "RCPT TO: " the first time it checked a socket. The thread stores this incomplete message. Then the second time around another thread picks up "example@example.com". The thread assembles the message into "RCPT TO: example@example.com" and then processes the entire command accordingly.
Does this model work better? Keep in mind that when DB calls need to be made, the MySQL library won't work the same way. A slow database server could hang all 8 worker threads effectively killing the model. There are also SPF, SPAM and Virus libraries. Any one of them could tie up a thread for an extended period, thereby killing this model. What does everyone else think? Am I not thinking about an event processing model correctly? Or is that this type of daemon is better off served using the one thread per connection model?
My company designed high-performance mono-process servers (portable ones too) starting in 1995, using event-driven virtual threads and state-machine frameworks. Very elegant, very fast, and really easy programming. The Xitami web server was one example - I remember seeing a Win95 system with Xitami survive a slashdotting (it was serving static pages but that was still impressive.)
We worked in C, because we needed guaranteed low latencies.
In 2004 we decided to rebuild these frameworks to handle OS multithreading. The reason was that on a single CPU we could not get the performance we needed, and the choice was either to use clusters, or multithreading.
We continued to work in C. C, and C++ are really nasty for multithreading because the languages have zero support for concurrency. You need to handle everything yourself, and most threading errors are extremely hard to detect.
It cost us about 10 times more to write our software as multithreaded code than using virtualised threads and we had to build whole reference management frameworks to ensure that threads could share data safely.
We did keep virtual threading, in fact, but virtual threads get handled by a pool of OS threads. Using 1 OS thread per connection is not scalable beyond a few hundred threads. Modern Linux kernels handle lots of threads but we also target Solaris, and Windows with the same code. So we use two virtual threads per connection, for full-duplex traffic, and we design most of the major server components as threaded objects, which are asynchronous event-driven objects.
Doing multithreading in C is a *huge* work. C++ has frameworks like ACE that help a lot.
But there is a performance gain. Our software is a messaging server (implementing the AMQP draft standard). We maxed out at around 55,000 messages per second using a pure virtual-threaded model. Very efficient code. On a single CPU the multithreaded code hits 35,000 messages per second. With two CPUs we're back at 55k, and with 4 dual-core Opterons we're at 120k-150k and higher. (Our software runs a massive trading application that processes 1.5bn messages per day). We still need to improve some of the low-level locking functions to use lock-free mechanisms, and we max out a gigabit network. It is difficult to find machines powerful enough to really stress test the software.
Without very robust frameworks, I'd never attempt such a project. As it was, we paid a lot for the extra performance. Our frameworks will eventually be released as free software, along with the middleware server.
Interestingly, a very similar application written in Java 1.5 and using the BEA runtime gets similar performance to ours. Java's threading is so good that I'd be hesitant to chose C on the basis of performance again. I'm not sure whether ACE can reach the levels of performance we need; 100k messages per second is extreme.
Other questions that are very important to ask:
- The number of clients you expect to connect at once. If it's less than 500 you can probably use one or two OS threads per connection. If it's more you need to virtualise connections or share your OS threads.
- The footprint. If you don't care, then I'd advise using Java. If you want a native Linux service, consider C++ and ACE. If you really want to write multithreaded C code, and don't have a full toolkit, consider seeing a doctor.
When it comes to the future, clearly multiple cores are the way we're heading. This was clear two years ago, and was the main reason we bit the bullet and chose to write our software multithreaded rather than using a clustering model. It seemed clear to me that within a decade, systems would have 32, 64, 128 cores, and software that could take advantage of this would survive for longer. Clustering is not as powerful an abstraction as multithreading.
My blog