I have a small communication problem that has consumed hours of search. I am using MPICH2 to communicate between different workers. At some points in my program a process needs to multi-cast a message to a fraction of the workers (2 or 3 out of a total of 20). Therefore, I temporarily need to create a group that includes the ranks of all those workers and then use MPI_BCast. However, this seems to be impossible!
I have tried MPI_Comm_Create but the program simply hangs because it required "every" worker call MPI_Comm_Create. I can not also use MPI_Comm_Split because I do not know the ranks of the recipient workers in advance and hence can not color code them.
Could you please help me.
Why do you need to create a new communicator at all?
Your description, of what you actually want to achieve and what the constraints are is a little lacking, but here are some hints, that might be applicable for your problem.
Sticking to classical two-sided communication, you need at some point a communication that involves all processes to identify the recipients, I guess. You could for example broadcast to everybody who is to be a recipient, and subsequently send the actual message to those with peer-to-peer communication (If this relation is going to change over time, I would not bother with creating a new communicator each time).
You could use MPI's one-sided communication concepts, and simply write messages from the broadcasting rank into dedicated memory areas of the receiving ranks. However, one-sided is often considered somewhat bad and not so good on the performance side.
With MPI-3 you could make use of an non-blocking barrier: All processes open the barrier, and those, which are not the broadcasting rank start immediately testing for the completion of this barrier, open a non-blocking receive for any source and regularly test for that as well, otherwise they proceed as usual. The broadcasting rank however, starts sending out its message to the actual recipients and when it completed that, it waits for the non-blocking barrier to complete. Now, all processes will find the barrier to complete, and now they can stop listening for the receives, those who didn't get a message can simply send a message to themselves to properly close the communication and proceed in their computation.
Related
This may appear as a silly question, but I am really confused about the terminology of the ZeroMQ regarding synchronous sockets like REQ and REP.
By my understanding a synchronous communication occurs when a client sends a message an then it blocks, until the response arrives. If ZeroMQ implemented a synchronous communication then only a .send() method would be enough for a synchronous socket.
I think that synchronous sockets terminology of ZeroMQ refers only to the inability of sending more messages until the response of the last message arrives, but the "sender" can still continue its processing ( doing more stuff ) asynchronously.
Is this true?
In that case, is there any straightforward way to implement a synchronous communication using ZeroMQ?
EDIT: Synchronous communication makes sense when I want to invoke a method in a remote process (like RPC). If I want to execute a series of commands in a remote process and each command needs the result of the previous one to do its job then asynchronous communication is not the best option.
To use ZMQ for implementing a synchronous framework, you can very nearly do it using just ZMQ; you can set the high water mark to 1. Unfortunately that's not quite it; what you want is an out going queue length of 0. Even more unfortunately, setting the high water mark to 0 is interpretted by ZMQ as infinity...
So the only option is to implement a synchronous transfer protocol on top of ZMQ. That's not very difficult to do. The conversation between the two ends will be something like "can I send?", "yes you can send now", "ok here it is", "ok I have received it" (both ends return to caller) (or at least the programatic version of that). This sets up what is called an execution rendevous - both ends know that they both reached a certain point of execution.
Technically speaking what you're doing is taking ZeroMQ (Actor Model) and turning it into something more like Communicating Sequential Processes.
RPC
Having said all that, from your edit I think you might like to consider Cap'n Proto. This is a C++ serialisation technology that has a neat RPC trick. If the return from one RPC call is the input to another, you can chain those all together somehow in advance (see here).
Let's start with a first stepforget everything you know about sockets.
ZeroMQ is more a concept of thinking about distributed-systems ( multi-agent like ) and how to design a software, with a use of such a smart signalling / messaging framework.
This is the core aim of the ZeroMQ, to allow designers remain thinking in the application domain and let all the low level dirty work to be operated actually without much of the designers' need to care of.
If have just recently started with ZeroMQ, one may enjoy a short read about a ZeroMQ global view first, before discussing details.
Having read and understood the concept of the ZeroMQ hierarchy, it is way simpler to start on details:
given a local Context() instance is a data-pumping engine and having a REQ/REP Scalable Formal Communications Archetype pattern in mind, the story is now actually a story about a network of distributed-Finite-State-Automata.
local process, operating just one side of the distributed REQ/REP communication archetype has zero power to influence the remote process to receive or not the message that was passed from the local process over to the ZeroMQ delivery services towards the indended recipient(s) in a fair belief. The less the local process can influence the remote process' intent to respond at all or not, so welcome to the realms of distributed multi-agent games.
Both the REQ and the REP formal behaviour has to meet its both the { local | distributed-mode }-expected sort of behaviour -- REQ asks first, REP answers then, so as to keep the contracted promise. The point is, that this behaviour is distributed and split among a pair of nodes, plus there are cases, when network incidents may throw the distributed-FSA into an unsalvageable mutual deadlock ( one may find more posts on this here zeromq quite often ).
So, your local-side REQ code imperatively .send()-s and has no obligation to stop without doing anything reasonable until REP-side .recv( zmq.NOBLOCK )-s or not ( no one has any kind of warranty a remote node exists at all, similarly, one has to set oneselves ready to anticipate and handle all cases, where a remote side will never respond, so many "new" challenges appear from the nature of a distributed multi-agent ecosystem ).
There are smart ways to handle this new breed of distributed chaos and uncertainties, using, best using .poll() and non-blocking forms of either the .send() and .recv()-methods, as these let user-code to remain capable of handling all expected and un-expected events in due time and fashion.
One may also operate rather many co-existent ZeroMQ connections, so as to prioritise and specialise each and any form of the multi-agents' interactions in a distributed system design, even for designing in fault-resilience and similar high-level robustness concept, where asynchronous nature of each of the interactions avoids a need of any sort of coordination or synchronisation with a remote ( possibly even not yet present ) agent, which is principally an autonomous entity, having it's own domain of control, so again, being principally asynchronous to what local-side agent might "expect", the less "influence" in any other form but by an attempt to send "there" a message "telegram".
So yes,ZeroMQ is asynchronous brokerless signalling / messaging framework.
For (almost) synchronous communications, one may take steps and measures to trim down the ( principally distributed ) asynchronous control loops -- best update your post with an MCVE example and details about what are your particular goals for being achieved.
I want the speed of asynchronous messages but still have some flow control. How can I accomplish this in Erlang?
There is no process memory limit right now -- it is discussed on mailing list etc. You can look at those threads.
On the up side, when you use OTP patterns implementation like gen_server you have a lot of freedom in retrieving messages from process queue and measuring the length of the queue.
gen_server2 used in rabbitmq used to optimize that by moving messages to internal data structure.
Having that you can discard any new incoming message when internal queue is too long.
You can do it silently or notify sender that the message rejected.
All of that is on very low level.
RabbitMQ will provide this functionality on AMQP level.
A common and quite good way of enforcing flow control is to make well selected messages into calls which limits how much load each client can load the server to one, effectively providing force feed back in an extremely simple way. The trick is of course to pick which communications uses synchronous calls :-)
I have a setup in which two nodes are going to be communicating a lot. On Node A, there are going to be thousands of processes, which are meant to access services on Node B. There is going to be a massive load of requests and responses across the two nodes. The two Nodes, will be running on two different servers, each on its own hardware server.
I have 3 Options: HTTP/1.1 , rpc:call/4 and Directly sending a message to a registered gen_server on Node B. Let me explain each option.
HTTP/1.1 Suppose that on Node A, i have an HTTP Client like Ibrowse, and on Node B, i have a web server like Yaws-1.95, the web server being able to handle unlimited connections, the operating system settings tweaked to allow yaws to handle all connections. And then make my processes on Node A to communicate using HTTP. In this case each method call, would mean a single HTTP request and a reply. I believe there is an overhead here, but we are evaluating options here. The erlang Built in mechanism called webtool, may be built for this kind of purpose.
rpc:call/4 I could simply make direct rpc calls from Node A to Node B. I am not very susre how the underlying rpc mechanism works , but i think that when two erlang nodes connect via net_adm:ping/1, the created connection is not closed but all rpc calls use this pipe to transmit requests and pass responses. Please correct me on this one.Sending a Message from Node A to Node B I could make my processes on Node A to just send message to a registered process, or a group of processes on Node B. This too seems a clean option.
Q1. Which of the above options would you recommend and why, for an application in which two erlang nodes are going to have enormous communications between them all the time. Imagine a messaging system, in which two erlang nodes are the routers :) ? Q2. Which of the above methods is cleaner, less problematic and is more fault tolerant (i mean here that, the method should NOT have single point of failure, that could lead to all processes on Node A blind) ? Q3. The mechanism of your choice: how would you make it even more fault tolerant, or redundant? Assumptions: The Nodes are always alive and will never go down, the network connection between the nodes will always be available and non-congested (dedicated to the two nodes only) , the operating system have allocated maximum resources to these two nodes. Thank you for your evaluations
HTTP is definitely out. Just the round-trip overhead of creating a new connection is a problem.
As for Erlang connections and using Pids, you have the advantage that you can subscribe to node-down messages and handle the case where a node goes down. A single TCP connection should be able to give you very fast speeds, however, be aware that it works like a long pipe: messages are muxed and demuxed on a pipe which can affect latency on the line. It also means that large messages will block small messages from getting through.
How much bandwidth are you aiming for, and at what latency? What is the 95th and 99th percentile of answering messages? It is better to put up some rough numbers and then try to target these than just having "as fast as possible". Set your success criteria first.
Q1: HTTP will add additional overhead and will give you nothing in my opinion. HTTP would be useful if you were designing a REST API. Directly sending messages and rpc:call look about the same as far as overhead is regarded.
Q2: Sending messages is much much clearer. It's the way erlang is designed. With RPC calls you must always track which call is executed where and under which circumstances which can be a huge issue if the two servers have state. Also RPC calls are synchronous.
Q3: I would use UBF if I can afford minor overhead, otherwise I would directly send messages between the erlang nodes. If the bandwidth is an issue other trickery would be needed as well. Like encoding the messages in same way and then using some compression algorithm to reduce the size of the message, alternatively I may ditch the erlang message passing altogether and use UDP sockets.
It is not obvious that ! is the best way to go. Definitely, it is the easiest and the code will be the most elegant.
In terms of scalability, take under consideration that to use rpc/! you have to maintain an erlang cluster. I found it painful having just 10-20 nodes even in private cloud. I would never recommend bigger deployments on e.g. EC2, where io/latency/network is not deterministic.
I recommend to structure the project in a way that will let you exchange communication engine in the future. Also HTTP is pretty heavy, but there are options:
socket-socket (tcp/udp/sctp)
amqp (many benefits connected to load balancing)
zeromq (even nicer than amqp)
Betting on !/rpc and OTP cluster is risky. You will fight with full mesh overhead, master election algos and quorum/partition detection.
I would like to speedup my MPI- Program with the use of asynchronous communication. But the used time remains the same. The workflow is as followed.
before:
1. MPI_send/ MPI_recv Halo (ca. 10 Seconds)
2. process the whole Array (ca. 12 Seconds)
after:
1. MPI_Isend/ MPI_Irecv Halo (ca. 0,1 Seconds)
2. process the Array (without Halo) (ca. 10 Seconds)
3. MPI_Wait (ca. 10 Seconds) (should be ca. 0 Seconds)
4. process the Halo only (ca. 2 Seconds)
Measurements showed that the communication and processing the Array-core nearly take the same time for common workloads. So asynchronism should nearly hide the communication time.
But it dosn't.
One fact - and I thinks this could be the problem - is that the sendbuffer is also the array the calculations are made on. Is it possible that MPI serializes the memory-access although communication ONLY accesses the Halo (with derived datatype) and the computation ONLY accesses the core (only reading) of the array???
Does anybody know if this is for sure the reason?
Is it maybe implementation-dependend (I'm using OpenMPI)?
Thanks in advance.
It isn't the case that MPI serializes the memory accesses in the user code (that's beyond the library's power to do, in general), and it is true that what exactly does happen is implementation specific.
But as a practical matter, MPI libraries don't do as much communication "in the background" as you might hope, and this is particularly true when using transports and networks like tcp + ethernet, where there's no meaningful way to hand off communication to another set of hardware.
You can only be sure that the MPI library is actually doing something when you're running MPI library code, eg in an MPI function call. Often, a call to any of a number of MPI calls will nudge an implementations "progress engine" that keeps track of in-flight messages and ushers them along. So for instance one thing you can quickly do is to make calls to MPI_Test() on the requests within the compute loop to make sure things start happening well before the MPI_Wait(). There is of course overhead to this, but this is something that's easy to try to measure.
Of course you could imagine the MPI library would use some other mechanism to run things behind the scenes. Both MPICH2 and OpenMPI have played with separate "progress threads" which execute separately from the user code and do this ushering along in the background; but getting that to work well, and without tying up a processor while you're trying to run your computation, is a genuinely difficult problem. OpenMPI's progress threads implementation has long been experimental, and in fact is temporarily out of the current (1.6.x) release, although work continues. I'm not sure about MPICH2's support.
If you are using infiniband, where the network hardware has a lot of intelligence to it, then prospects brighten a bit. If you are willing to leave memory pinned (for the openfabrics), and/or you can use a vendor-specific module (mxm for Mellanox, psm for Qlogic), then things can progress somewhat more rapidly. If you're using shared memory, than the knem kernel module can also help with intranode transport.
One other implementation-specific approach you can take, if memory isn't a big issue, is to try to use eager protocols for sending the data directly, or send more data per chunk so fewer nudges of the progress engine are needed. What eager protocols means here is that data is automatically sent at send time, rather than just initiating a set of handshakes which will eventually lead to the message being sent. The bad news is that this generally requires extra buffer memory for the library, but if that's not a problem and you know the number of incoming messages is bounded (eg, by the number of halo neighbours you have), this can help a great deal. How to do this for (eg) shared memory transport for openmpi is described on the OpenMPI page for tuning for shared memory, but similar parameters exist for other transports and often for other implementations. One nice tool that IntelMPI has is an "mpitune" tool that automatically runs through a number of such parameters for best performance.
The MPI specification states:
A nonblocking send call indicates that the system may start copying
data out of the send buffer. The sender should not modify any part of the
send buffer after a nonblocking send operation is called, until the
send completes.
So yes, you should copy your data to a dedicated send buffer first.
This question is the result of two other questions I've asked in the last few days.
I'm creating a new question because I think it's related to the "next step" in my understanding of how to control the flow of my send/receive, something I didn't get a full answer to yet.
The other related questions are:
An IOCP documentation interpretation question - buffer ownership ambiguity
Non-blocking TCP buffer issues
In summary, I'm using Windows I/O Completion Ports.
I have several threads that process notifications from the completion port.
I believe the question is platform-independent and would have the same answer as if to do the same thing on a *nix, *BSD, Solaris system.
So, I need to have my own flow control system. Fine.
So I send send and send, a lot. How do I know when to start queueing the sends, as the receiver side is limited to X amount?
Let's take an example (closest thing to my question): FTP protocol.
I have two servers; One is on a 100Mb link and the other is on a 10Mb link.
I order the 100Mb one to send to the other one (the 10Mb linked one) a 1GB file. It finishes with an average transfer rate of 1.25MB/s.
How did the sender (the 100Mb linked one) knew when to hold the sending, so the slower one wouldn't be flooded? (In this case the "to-be-sent" queue is the actual file on the hard-disk).
Another way to ask this:
Can I get a "hold-your-sendings" notification from the remote side? Is it built-in in TCP or the so called "reliable network protocol" needs me to do so?
I could of course limit my sendings to a fixed number of bytes but that simply doesn't sound right to me.
Again, I have a loop with many sends to a remote server, and at some point, within that loop I'll have to determine if I should queue that send or I can pass it on to the transport layer (TCP).
How do I do that? What would you do? Of course that when I get a completion notification from IOCP that the send was done I'll issue other pending sends, that's clear.
Another design question related to this:
Since I am to use a custom buffers with a send queue, and these buffers are being freed to be reused (thus not using the "delete" keyword) when a "send-done" notification has been arrived, I'll have to use a mutual exlusion on that buffer pool.
Using a mutex slows things down, so I've been thinking; Why not have each thread have its own buffers pool, thus accessing it , at least when getting the required buffers for a send operation, will require no mutex, because it belongs to that thread only.
The buffers pool is located at the thread local storage (TLS) level.
No mutual pool implies no lock needed, implies faster operations BUT also implies more memory used by the app, because even if one thread already allocated 1000 buffers, the other one that is sending right now and need 1000 buffers to send something will need to allocated these to its own.
Another issue:
Say I have buffers A, B, C in the "to-be-sent" queue.
Then I get a completion notification that tells me that the receiver got 10 out of 15 bytes. Should I re-send from the relative offset of the buffer, or will TCP handle it for me, i.e complete the sending? And if I should, can I be assured that this buffer is the "next-to-be-sent" one in the queue or could it be buffer B for example?
This is a long question and I hope none got hurt (:
I'd loveeee to see someone takes the time to answer here. I promise I'll double-vote for him! (:
Thank you all!
Firstly: I'd ask this as separate questions. You're more likely to get answers that way.
I've spoken about most of this on my blog: http://www.lenholgate.com but then since you've already emailed me to say that you read my blog you know that...
The TCP flow control issue is such that since you are posting asynchronous writes and these each use resources until they complete (see here). During the time that the write is pending there are various resource usage issues to be aware of and the use of your data buffer is the least important of them; you'll also use up some non-paged pool which is a finite resource (though there is much more available in Vista and later than previous operating systems), you'll also be locking pages in memory for the duration of the write and there's a limit to the total number of pages that the OS can lock. Note that both the non-paged pool usage and page locking issues aren't something that's documented very well anywhere, but you'll start seeing writes fail with ENOBUFS once you hit them.
Due to these issues it's not wise to have an uncontrolled number of writes pending. If you are sending a large amount of data and you have a no application level flow control then you need to be aware that if you send data faster than it can be processed by the other end of the connection, or faster than the link speed, then you will begin to use up lots and lots of the above resources as your writes take longer to complete due to TCP flow control and windowing issues. You don't get these problems with blocking socket code as the write calls simply block when the TCP stack can't write any more due to flow control issues; with async writes the writes complete and are then pending. With blocking code the blocking deals with your flow control for you; with async writes you could continue to loop and more and more data which is all just waiting to be sent by the TCP stack...
Anyway, because of this, with async I/O on Windows you should ALWAYS have some form of explicit flow control. So, you either add application level flow control to your protocol, using an ACK, perhaps, so that you know when the data has reached the other side and only allow a certain amount to be outstanding at any one time OR if you cant add to the application level protocol, you can drive things by using your write completions. The trick is to allow a certain number of outstanding write completions per connection and to queue the data (or just don't generate it) once you have reached your limit. Then as each write completes you can generate a new write....
Your question about pooling the data buffers is, IMHO, premature optimisation on your part right now. Get to the point where your system is working properly and you have profiled your system and found that the contention on your buffer pool is the most important hot spot and THEN address it. I found that per thread buffer pools didn't work so well as the distribution of allocations and frees across threads tends not to be as balanced as you'd need to that to work. I've spoken about this more on my blog: http://www.lenholgate.com/blog/2010/05/performance-comparisons-for-recent-code-changes.html
Your question about partial write completions (you send 100 bytes and the completion comes back and says that you have only sent 95) isn't really a problem in practice IMHO. If you get to this position and have more than the one outstanding write then there's nothing you can do, the subsequent writes may well work and you'll have bytes missing from what you expected to send; BUT a) I've never seen this happen unless you have already hit the resource problems that I detail above and b) there's nothing you can do if you have already posted more writes on that connection so simply abort the connection - note that this is why I always profile my networking systems on the hardware that they will run on and I tend to place limits in MY code to prevent the OS resource limits ever being reached (bad drivers on pre Vista operating systems often blue screen the box if they can't get non paged pool so you can bring a box down if you don't pay careful attention to these details).
Separate questions next time, please.
Q1. Most APIs will give you "write is possible" event, after you last wrote and writing is available again (can happen immediately if you failed to fill major part of send buffer with the last send).
With completion port, it will arrive just as "new data" event. Think of new data as "read Ok", so there's also a "write ok" event. Names differ between the APIs.
Q2. If a kernel mode transition for mutex acquisition per chunk of data hurts you, I recommend rethinking what you are doing. It takes 3 microseconds at most, while your thread scheduler slice may be as big as 60 milliseconds on windows.
It may hurt in extreme cases. If you think you are programming extreme communications, please ask again, and I promise to tell you all about it.
To address your question about when it knew to slow down, you seem to lack an understanding of TCP congestion mechanisms. "Slow start" is what you're talking about, but it's not quite how you've worded it. Slow start is exactly that -- starts off slow, and gets faster, up to as fast as the other end is willing to go, wire line speed, whatever.
With respect to the rest of your question, Pavel's answer should suffice.