What is the role of 'cluster' software in relation to MPI? - c

I'm a little confused regarding how a cluster implementation ("Beowulf cluster") relates to a communication protocol such as MPI. What software components are needed to set up a "cluster" using something like OpenMPI?

A cluster, as you know, is a bunch of computers networked together. When you have such configuration, you normally install and use the following:
MPI, for communication between processes
NFS, to have a network disk visible and shared to all nodes
NTP, to synchronize the time of the nodes so that you can compare log events and timestamps
bootp to boot the nodes from a remote node, so that each node restart fresh with a guaranteed good and uniform setup.
a set of cluster utilities to make your life easier, such as a distributed ssh to execute the same command on all nodes at the same time.
a task scheduler, or queue manager, such as Condor, LFS or others, that allow you to prioritize job submissions and eventually measure them for limiting/pricing.
a watchdog, so to reboot one node automatically if it gets stuck.
software control for UPS (so to shut down automatically in case of prolonged loss of power)
And much more. All this stuff is completely additional to MPI. MPI is just a communication channel between processes. MPI alone does not "make the cluster".

MPI, as you noted, will only provide communication between processes. If there will not be several people using the cluster, you really need nothing more (apart from some script to launch your program on all the nodes).
But, in reality we sadly seldom have our personal cluster. That's when you need a scheduler. The scheduler typically handles job submissions and resource allocation, possibly also taking care of prioritization, user management and other things to make your life easier.

Take a look at Oracle Grid Engine (nee Sun Grid Engine or CODINE).

Related

Erlang spawning large amounts of C processes

I've been looking into how I could embed languages (let's use Lua as an example) in Erlang. This of course isn't a new idea and there are many libraries out there that can do this. However I was wondering if it was possible to start a Genserver with state which is modified by Lua. This means that once you start the Genserver, it will start a (long running) Lua process to manipulate the Genserver's state. I know this is possible as well, but I was wondering if I could spawn 1,000 10,000 or even 100,000 of these processes.
I'm not really familiar with this topic but I have done some research.
(Please correct me if I'm wrong on any of these options).
TLDR; Skip to the last paragraph.
First option: NIFs:
This doesn't seem like an option since it will block the Erlang Scheduler of the current process. If I want to spawn a large amount of these it will freeze the entire runtime.
Second option: Port Driver:
It's like a NIF but communicates by sending data to a specified port, which can also send data back to Erlang. This is nice although this also seems to block the scheduler. I've tried a library which does the boiler plat for you as well, but that seemed to block the scheduler after spawning 10 processes. I've also looked into the postgresql example on the Erlang Documentation which is said to be async but I couldn't get the example code to work (R13?). Is it even possible to run as many Port Driver processes without blocking the runtime?
Third option: C Nodes:
I thought this was very interesting and wanted to try it out, but apparently the project "erlang-lua" already does this. It's nice because it won't crash your Erlang VM if something goes wrong and the processes are isolated. But in order to actually spawn a single process you need to spawn an entire node. I have no idea how expensive this is. Nor am I sure what the limit is for connecting nodes in a cluster, but I don't see myself spawning 100,000 C nodes.
Fourth option: Ports:
At first I thought this was the same as a Port Driver but it's actually different. You spawn a process which executes an application and communicates through STDIN and STDOUT. This would work well for spawning a large amount of processes, and (I think?) they aren't a threat to the Erlang VM. But if I'm going to communicate through STDIN / STDOUT, why even bother with an embeddable language to begin with? Might as well use any other scripting language.
And so after much research in a field I'm not familiar with I've come to this. You could a Genserver as an "entity" where the AI is written in Lua. Which is why I'd like to have a processes for each entity. My question is how do I achieve spawning many Genservers which communicate with long running Lua processes? Is this even possible? Should I be tackling my problem differently?
If you can make the Lua code — or more accurately, its underlying native code — cooperate with the Erlang VM, you have a few choices.
Consider one of the most important functions of the Erlang VM: managing the execution of a (potentially large number of) Erlang's lightweight processes across a relatively small set of scheduler threads. It uses several techniques to know when a process has used up its timeslice or is waiting and so should be scheduled out to give another process a chance to run.
You seem to be asking how you can get native code to run however it likes within the VM, but as you've already hinted, the reason native code can cause problems for the VM is that it has no practical way to stop the native code from completely taking over a scheduler thread and thus preventing regular Erlang processes from executing. Because of this, native code has to cooperatively yield the scheduler thread back to the VM.
For older NIFs, the choices for such cooperation were:
Keep the amount of time NIF calls ran on a scheduler thread to 1ms or less.
Create one or more private threads. Transition each long-running NIF call from its scheduler thread over to a private thread for execution, then return the scheduler thread to the VM.
The problems here are that not all calls can complete in 1ms or less, and that managing private threads can be error-prone. To get around the first problem, some developers would break the work down into chunks and use an Erlang function as a wrapper to manage a series of short NIF calls, each of which completed one chunk of work. As for the second problem, well, sometimes you just can't avoid it, despite its inherent difficulty.
NIFs running on Erlang 17.3 or later can also cooperatively yield the scheduler thread using the enif_schedule_nif function. To use this feature, the native code has to be able to do its work in chunks such that each chunk can complete within the usual 1ms NIF execution window, similar to the approach mentioned earlier but without the need to artificially return to an Erlang wrapper. My bitwise example code provides many details about this.
Erlang 17 also brought an experimental feature, off by default, called dirty schedulers. This is a set of VM schedulers that do not have the same native code execution time constraints as the regular schedulers; work there can block for essentially infinite periods without disrupting normal VM operation.
Dirty schedulers come in two flavors: CPU schedulers for CPU-bound work, and I/O schedulers for I/O-bound work. In a VM compiled to enable dirty schedulers, there are by default as many dirty CPU schedulers as there are regular schedulers, and there are 10 I/O schedulers. These numbers can be altered using command-line switches, but note that to try to prevent regular scheduler starvation, you can never have more dirty CPU schedulers than regular schedulers. Applications use the same enif_schedule_nif function mentioned earlier to execute NIFs on dirty schedulers. My bitwise example code provides many details about this too. Dirty schedulers will remain an experimental feature for Erlang 18 as well.
Native code in linked-in port drivers is subject to the same on-scheduler execution time constraints as NIFs, but drivers have two features NIFs don't:
Driver code can register file descriptors into the VM polling subsystem and be notified when any of those file descriptors becomes I/O-ready.
The driver API supports access to a non-scheduler async thread pool, the size of which is configurable but by default has 10 threads.
The first feature allows native driver code to avoid blocking a thread for I/O. For example, instead of performing a blocking recv call, driver code can register the socket file descriptor so the VM can poll it and call the driver back when the file descriptor becomes readable.
The second feature provides a separate thread pool useful for driver tasks that can't conform to the scheduler thread native code execution time constraints. You can achieve the same in a NIF but you have to set up your own thread pool and write your own native code to manage and access it. But regardless of whether you use the driver async thread pool, your own NIF thread pool, or dirty schedulers, note that they are all regular operating system threads, and so trying to start a huge number of them simply isn't practical.
Native driver code does not yet have dirty scheduler access, but this work is on-going and it might become available as an experimental feature in an 18.x release.
If your Lua code can make use of one or more of these features to cooperate with the Erlang VM, then what you're attempting may be possible.

What is the best way to avoid overloading a parallel file-system when running embarrassingly parallel jobs?

We have a problem which is embarrassingly parallel - we run a large number of instances of a single program with a different data set for each; we do this simply by submitting the application many times to the batch queue with different parameters each time.
However with a large number of jobs, not all of them complete. It does not appear to be a problem in the queue - all of the jobs are started.
The issue appears to be that with a large number of instances of the application running, lots of jobs finish at roughly the same time and thus all try to write out their data to the parallel file-system at pretty much the same time.
The issue then seems to be that either the program is unable to write to the file-system and crashes in some manner, or just sits there waiting to write and the batch queue system kills the job after it's been sat waiting too long. (From what I have gathered on the problem, most of the jobs that fail to complete, if not all, do not leave core files)
What is the best way to schedule disk-writes to avoid this problem? I mention our program is embarrassingly parallel to highlight the fact the each process is not aware of the others - they cannot talk to each other to schedule their writes in some manner.
Although I have the source-code for the program, we'd like to solve the problem without having to modify this if possible as we don't maintain or develop it (plus most of the comments are in Italian).
I have had some thoughts on the matter:
Each job write to the local (scratch) disk of the node at first. We can then run another job which checks every now and then what jobs have completed and moves the files from the local disks to the parallel file-system.
Use an MPI wrapper around the program in master/slave system, where the master manages a queue of jobs and farms these off to each slave; and the slave wrapper runs the applications and catches the exception (could I do this reliably for a file-system timeout in C++, or possibly Java?), and sends a message back to the master to re-run the job
In the meantime I need to pester my supervisors for more information on the error itself - I've never run into it personally, but I haven't had to use the program for a very large number of datasets (yet).
In case it's useful: we run Solaris on our HPC system with the SGE (Sun GridEngine) batch queue system. The file-system is NFS4, and the storage servers also run Solaris. The HPC nodes and storage servers communicate over fibre channel links.
Most parallel file systems, particularly those at supercomputing centres, are targetted for HPC applications, rather than serial-farm type stuff. As a result, they're painstakingly optimized for bandwidth, not for IOPs (I/O operations per sec) - that is, they are aimed at big (1000+ process) jobs writing a handful of mammoth files, rather than zillions of little jobs outputting octillions of tiny little files. It is all to easy for users to run something that runs fine(ish) on their desktop and naively scale up to hundreds of simultaneous jobs to starve the system of IOPs, hanging their jobs and typically others on the same systems.
The main thing you can do here is aggregate, aggregate, aggregate. It would be best if you could tell us where you're running so we can get more information on the system. But some tried-and-true strategies:
If you are outputting many files per job, change your output strategy so that each job writes out one file which contains all the others. If you have local ramdisk, you can do something as simple as writing them to ramdisk, then tar-gzing them out to the real filesystem.
Write in binary, not in ascii. Big data never goes in ascii. Binary formats are ~10x faster to write, somewhat smaller, and you can write big chunks at a time rather than a few numbers in a loop, which leads to:
Big writes are better than little writes. Every IO operation is something the file system has to do. Make few, big, writes rather than looping over tiny writes.
Similarly, don't write in formats which require you to seek around to write in different parts of the file at different times. Seeks are slow and useless.
If you're running many jobs on a node, you can use the same ramdisk trick as above (or local disk) to tar up all the jobs' outputs and send them all out to the parallel file system at once.
The above suggestions will benefit the I/O performance of your code everywhere, not juston parallel file systems. IO is slow everywhere, and the more you can do in memory and the fewer actual IO operations you execute, the faster it will go. Some systems may be more sensitive than others, so you may not notice it so much on your laptop, but it will help.
Similarly, having fewer big files rather than many small files will speed up everything from directory listings to backups on your filesystem; it is good all around.
It is hard to decide if you don't know what exactly causes the crash. If you think it is an error related to the filesystem performance, you can try an distributed filesystem: http://hadoop.apache.org/common/docs/r0.20.0/hdfs_user_guide.html
If you want to implement Master/Slave system, maybe Hadoop can be the answer.
But first of all I would try to find out what causes the crash...
OSes don't alway behave nicely when they run out of resources; sometimes they simply abort the process that asks for the first unit of resource the OS can't provide. Many OSes have file handle resource limits (Windows I think has a several-thousand handle resource, which you can bump up against in circumstances like yours), and failure to find a free handle usually means the OS does bad things to the requesting process.
One simple solution requiring a program change, is to agree that no more than N of your many jobs can be writing at once. You'll need a shared semaphore that all jobs can see; most OSes will provide you with facilities for one, often as a named resource (!). Initialize the semaphore to N before you launch any job.
Have each writing job acquire a resource unit from the semaphore when the job is about to write, and release that resource unit when it is done. The amount of code to accomplish this should be a handful of lines inserted once into your highly parallel application. Then you tune N until you no longer have the problem. N==1 will surely solve it, and you can presumably do lots better than that.

Server Architecture for Embedded Device

I am working on a server application for an embedded ARM platform. The ARM board is connected to various digital IOs, ADCs, etc that the system will consistently poll. It is currently running a Linux kernel with the hardware interfaces developed as drivers. The idea is to have a client application which can connect to the embedded device and receive the sensory data as it is updated and issue commands to the device (shutdown sensor 1, restart sensor 2, etc). Assume the access to the sensory devices is done through typical ioctl.
Now my question relates to the design/architecture of this server application running on the embedded device. At first I was thinking to use something like libevent or libev, lightweight C event handling libraries. The application would prioritize the sensor polling event (and then send the information to the client after the polling is done) and process client commands as they are received (over a typical TCP socket). The server would typically have a single connection but may have up to a dozen or so, but not something like thousands of connections. Is this the best approach to designing something like this? Of the two event handling libraries I listed, is one better for embedded applications or are there any other alternatives?
The other approach under consideration is a multi-threaded application in which the sensor polling is done in a prioritized/blocking thread which reads the sensory data and each client connection is handled in separate thread. The sensory data is updated into some sort of buffer/data structure and the connection threads handle sending out the data to the client and processing client commands (I supposed you would still need an event loop of sort in these threads to monitor for incoming commands). Are there any libraries or typical packages used which facilitate designing an application like this or is this something you have to start from scratch?
How would you design what I am trying to accomplish?
I would use a unix domain socket -- and write the library myself, can't see any advantages to using libvent since the application is tied to linux, and libevent is also for hundreds of connections. You can do all of what you are trying to do with a single thread in your daemon. KISS.
You don't need a dedicated master thread for priority queues you just need to write your threads so that it always processes high priority events before anything else.
In terms of libraries, you will possibly benifit from Google's protocol buffers (for serialization and representing your protocol) -- however it only has first class supports for C++, and the over the wire (serialization) format does a bit of simple bit shifting to numeric data. I doubt it will add any serious overhead. However an alternative is ASN.1 (asn1c).
My suggestion would be a modified form of your 2nd proposal. I would create a server that has two threads. One thread polling the sensors, and another for ALL of your client connections. I have used in embedded devices (MIPS) boost::asio library with great results.
A single thread that handles all sockets connections asynchronously can usually handle the load easily (of course, it depends on how many clients you have). It would then serve the data it has on a shared buffer. To reduce the amount and complexity of mutexes, I would create two buffers, one 'active' and another 'inactive', and a flag to indicate the current active buffer. The polling thread would read data and put it in the inactive buffer. When it finished and had created a 'consistent' state, it would flip the flag and swap the active and inactive buffers. This could be done atomically and should therefore not require anything more complex than this.
This would all be very simple to set up since you would pretty much have only two threads that know nothing about the other.

Where can I find benchmarks on different networking architectures?

Where can I find benchmarks on different networking architectures?
I am playing with sockets / threads / forks and I'd like to know what the best is. I was thinking there has got to be a place where someone has already spelled out all the pros and cons of different architectures for a socket service, listed benchmarks with code that runs.
Ultimately I'd like to run these various configurations with my own code and see which runs best in different circumstances.
Many people I talk to say that I should just use single threaded select. But I see an argument for threads when you're storing state information inside the thread to keep code simple. What is the trade off mark for writing my own state structure vs using a proven thread architecture.
I've also been told forking is bad... but when you need 12000 connections on a machine that cannot raise the open file per process limit, forking is an option! Forking is also a nice option for stability when you've got one process that needs restarting, it doesn't disturb the others.
Sorry, this is one of my longer questions... so many variables are left empty.
Thanks,
Chenz
edit: here's the link I was looking for, which is a whole paper answering your question. http://www.kegel.com/c10k.html
There are web servers designed along all three models (fork, thread, select). People like to benchmark web servers.
http://www.lighttpd.net/benchmark
Libevent has some benchmarks and links to stuff about how to choose a select() vs. threaded model, generally in favour of using the libevent model.
http://monkey.org/~provos/libevent/
It's very difficult to answer this question as so much depends on what your service is actually doing. Does it have to query a database? read files from the filesystem? perform complicated calculations? go off and talk to some other service? Also, how long-lived are client connections? Might connections have some semantic interaction with other connections, or are they all treated as independent of each other? Might you want to think about load-balancing your service across multiple servers later? (If so, you might usefully think about that now so that any necessary help can be designed in from the start.)
As you hint, the serving machine might have limits which interact with the various techniques, steering you towards one answer or another. You have a per-process file descriptor limit, but remember that you may also have a fixed size process table! How many concurrent clients are you expecting, anyway?
If your service keeps crashing and you need to keep restarting it or you think you want a multi-process model so that connections are isolated from each other, you're probably doing it wrong. Stability is extremely important in this sort of context, and that means good practice and memory hygiene, both in general and in the face of network-based attacks.
Remember the history... fork() is cheap in the Unix world, but spawning new processes relatively expensive on Windows. OTOH, Windows threads are lightweight, whereas threading has always been a bit alien to Unix and only relatively recently become widespread.

How to scale a TCP listener on modern multicore/multisocket machines

I have a daemon to write in C, that will need to handle 20-150K TCP connections simultaneously. They are long running connections, and rarely ever tear down. They have a very small amount of data (rarely exceeding MTU even.. it's a stimulus/response protocol) in transmit at any given time, but response times to them are critical. I'm wondering what the current UNIX community is using to get large amounts of sockets, and minimizing the latency on response of them. I've seen designs revolving around multiplexing connects to fork worker pools, threads (per connection), static sized thread pools. Any suggestions?
the easiest suggestion is to use libevent, it makes it easy to write a simple non-blocking single-threaded server that would comply with your requirements.
if the processing for each response takes some time, or if it uses some blocking API (like almost anything from a DB), then you'll need some threading.
One answer is the worker threads, where you spawn a set of threads, each listening on some queue to work. it can be separate processes, instead of threads, if you like. The main difference would be the communications mechanism to tell the workers what to do.
A different way to do is to use several threads, and give to each of them a portion of those 150K connections. each will have it's own process loop and work mostly like the single-threaded server, except for the listening port, which will be handled by a single thread. This helps spreading the load between cores, but if you use a blocking resource, it would block all the connections handled by this specific thread.
libevent lets you use the second way if you're careful; but there's also an alternative: libev. it's not as well known as libevent, but it specifically supports the multi-loop scheme.
If performance is critical then you'll really want to go for a multithreaded event loop solution - i.e. a pool of worker threads to handle your connections. Unfortunately, there is no abstraction library to do this that works on most Unix platforms (note that libevent is only single-threaded as are most of these event-loop libraries), so you'll have to do the dirty work yourself.
On Linux that means using edge-triggered epoll with a pool of worker threads (Windows would have I/O completion ports which also works fine in a multithreaded environment - I am not sure about other Unixes).
BTW, I have done some work trying to abstract edge-triggered epoll on Linux and Windows I/O completion ports on http://nginetd.cmeerw.org (it is work in progress, but might provide some ideas).
If you have system configuration access don't over-do it and set up some iptables/pf/etc to load-balance connections across n daemon instances (processes) as this will work out of the box. Depending on how blocking the nature of the daemon n should be from the number of cores on the system or several times higher. This approach looks crude but it can handle broken daemons and even restart them if necessary. Also migration would be smooth as you could start diverting new connections to another set of processes (for example, a new release or migrating to a new box) instead of service interruptions. On top of that you get several features like source affinity wich can help significantly caching and contention of problematic sessions.
If you don't have system access (or ops can't be bothered), you can use load balancer daemon (there are plenty of open source ones) instead of iptables/pf/etc and use also n service daemons, like above.
Also this approach helps with separating privileges of ports. If the external service needs to service on a low port (<1024) you only need the load balancer running privileged/or admin/root, or kernel.)
I've written several IP load balancers in the past and it can be very error-prone in production. You don't want to support and debug that. Also operations and management will tend second-guess your code more than external code.
i think javier's answer makes the most sense. if you want to test the theory out, then check out the node javascript project.
Node is based on Google's v8 engine which compiles javascript to machine code and is as fast as c for certain tasks. It is also based on libev and is designed to be completely non-blocking, meaning you don't have to worry about context switching between threads (everything runs on a single event loop). It is very similar to erlang in that respect.
Writing high performance servers in javascript is now really, really easy with node. You could also, with a little bit of effort, write your custom code in c and create bindings for node to call into it to do your actual processing (look at the node source to see how to do this - documentation is a little sketchy at the moment). as an uglier alternative, you could build your custom c code as an application and use stdin/stdout to communicate with it.
I've tested node myself with upwards of 150k connections with absolutely no issues (of course you will need some serious hardware if all these connections are going to be communicating at once). A TCP connection in node.js on average uses only 2-3k of memory so you could theoretically handle 350-500k connections per 1GB of RAM.
Note - Node.js is not currently supported on windows, but it is only at an early stage of development and i'd imagine it will be ported at some stage.
Note 2 - you will have to ensure the code you are calling into from Node does not block
Several systems have been developed to improve on select(2) performance: kqueue, epoll, and /dev/poll. In all these systems, you can have a pool of worker threads waiting for tasks; you will not be forced to setup all file handles over and over again when done with one of them.
do you have to start from scratch? You could use something like gearman.

Resources