I have a single producer and n workers that I only want to give work to when they're not already processing a unit of work and I'm struggling to find a good zeroMQ pattern.
1) REQ/REP
The producer is the requestor and creates a connection to each worker. It tracks which worker is busy and round-robins to idle workers
Problem:
How to be notified of responses and still able to send new work to idle workers without dedicating a thread in the producer to each worker?
2) PUSH/PULL
Producer pushes into one socket that all workers feed off, and workers push into another socket that the producer listens to.
Problem:
Has no concept of worker idleness, i.e. work gets stuck behind long units of work
3) PUB/SUB
Non-starter, since there is no way to make sure work doesn't get lost
4) Reverse REQ/REP
Each worker is the REQ end and requests work from the producer and then sends another request when it completes the work
Problem:
Producer has to block on a request for work until there is work (since each recv has to be paired with a send ). This prevents workers to respond with work completion
Could be fixed with a separate completion channel, but the producer still needs some polling mechanism to detect new work and stay on the same thread.
5) PAIR per worker
Each worker has its own PAIR connection allowing independent sending of work and receipt of results
Problem:
Same problem as REQ/REP with requiring a thread per worker
As much as zeroMQ is non-blocking/async under the hood, I cannot find a pattern that allows my code to be asynchronous as well, rather than blocking in many many dedicated threads or polling spin-loops in fewer. Is this just not a good use case for zeroMQ?
Your problem is solved with the Load Balancing Pattern in the ZMQ Guide. It's all about flow control whilst also being able to send and receive messages. The producer will only send work requests to idle workers, whilst the workers are able to send and receive other messages at all times, e.g. abort, shutdown, etc.
Push/Pull is your answer.
When you send a message in ZeroMQ, all that happens initially is that it sits in a queue waiting to be delivered to the destination(s). When it has been successfully transferred it is removed from the queue. The queue is limited in length, but can be set by changing a socket's high water mark.
There is a/some background thread(s) that manage all this on your behalf, and your calls to the ZeroMQ API are simply issuing instructions to that/those threads. The threads at either end of a socket connection are collaborating to marshall the transfer of messages, i.e. a sender won't send a message unless the recipient can receive it.
Consider what this means in a push/pull set up. Suppose one of your pull workers is falling behind. It won't then be accepting messages. That means that messages being sent to it start piling up until the highwater mark is reached. ZeroMQ will no longer send messages to that pull worker. In fact AFAIK in ZeroMQ, a pull worker whose queue is more full than those of its peers will receive less messages, so the workload is evened out across all workers.
So What Does That Mean?
Just send the messages. Let 0MQ sort it out for you.
Whilst there's no explicit flag saying 'already busy', if messages can be sent at all then that means that some pull worker somewhere is able to receive it solely because it has kept up with the workload. It will therefore be best placed to process new messages.
There are limitations. If all the workers are full up then no messages are sent and you get blocked in the push when it tries to send another message. You can discover this only (it seems) by timing how long the zmq_send() took.
Don't Forget the Network
There's also the matter of network bandwidth to consider. Messages queued in the push will tranfer at the rate at which they're consumed by the recipients, or at the speed of the network (whichever is slower). If your network is fundamentally too slow, then it's the Wrong Network for the job.
Latency
Of course, messages piling up in buffers represents latency. This can be restricted by setting the high water mark to be quite low.
This won't cure a high latency problem, but it will allow you to find out that you have one. If you have an inadequate number of pull workers, a low high water mark will result in message sending failing/blocking sooner.
Actually I think in ZeroMQ it blocks for push/pull; you'd have to measure elapsed time in the call to zmq_send() to discover whether things had got bottled up.
Thought about Nanomsg?
Nanomsg is a reboot of ZeroMQ, one of the same guys is involved. There's many things I prefer about it, and ultimately I think it will replace ZeroMQ. It has some fancier patterns which are more universally usable (PAIR works on all transports, unlike in ZeroMQ). Also the patterns are essentially a plugable component in the source code, so it is far simpler for patterns to be developed and integrated than in ZeroMQ. There is a discussion on the differences here
Philisophical Discussion
Actor Model
ZeroMQ is definitely in the realms of Actor Model programming. Messages get stuffed into queues / channels / sockets, and at some undetermined point in time later they emerge at the recipient end to be processed.
The danger of this type of architecture is that it is possible to have the potential for deadlock without knowing it.
Suppose you have a system where messages pass both ways down a chain of processes, say instructions in one way and results in the other. It is possible that one of the processes will be trying to send a message whilst the recipient is actually also trying to send a message back to it.
That only works so long as the queues aren't full and can (temporarily) absorb the messages, allowing everyone to move on.
But suppose the network briefly became a little busy for some reason, and that delayed message transfer. The message send might then fail because the high water mark had been reached. Whoops! No one is then sending anything to anyone anymore!
CSP
A development of the Actor Model, called Communicating Sequential Processes, was invented to solve this problem. It has a restriction; there is no buffering of messages at all. No process can complete sending a message until the recipient has received all the data.
The theoretical consequence of this was that it was then possible to mathematically analyse a system design and pronounce it to be free of deadlock. The practical consequence is that if you've built a system that can deadlock, it will do so every time. That's actually not so bad; it'll show up in testing, not post-deployment.
Curiously this is hinted at in the documentation of Microsoft's Task Parallel library, where they advocate setting buffer lengths to zero in the intersts of achieving a more robust application.
It'd be like setting the ZeroMQ high water mark to zero, but in zmq_setsockopt() 0 means default, not nought. The default is non-zero...
CSP is much more suited to real time applications. Any shortage of available workers immediately results in an inability to send messages (so your system knows it's failed to keep up with the real time demand) instead of resulting in an increased latency as data is absorbed by sockets, etc. (which is far harder to discover).
Unfortunately almost every communications technology we have (Ethernet, TCP/IP, ZeroMQ, nanomsg, etc) leans towards Actor Model. Everything has some sort of buffer somewhere, be it a packet buffer on a NIC or a socket buffer in an operating system.
Thus to implement CSP in the real world one has to implement flow control on top of the existing transports. This takes work, and it's slightly inefficient. But if a system that needs it, it's definitely the way to go.
Personally I'd love to see 0MQ and Nanomsg to adopt it as a behavioural option.
Related
There is a microservice, which receives the batch of the messages from the outside and push them to kafka. Each message is sent separately, so for each batch I have around 1000 messages 100 bytes each. It seems like the messages take much more space internally, because the free space on the disk going down much faster than I expected.
I'm thinking about changing the producer logic, the way it will put all the batch in one message (the consumer then will split them by itself). But I haven't found any information about space or performance issues with many small messages, neither any guildlines about balance between size and count. And I don't know Kafka enough to have my own conclusion.
Thank you.
The producer will, by itself, batch messages that are destined to the same partition, in order to avoid unnecesary calls.
The producer makes this thanks to its background threads. In the image, you can see how it batches 3 messages before sending them to each partition.
If you also set compression in the producer-side, it will also compress (GZip, LZ4, Snappy are the valid codecs) the messages before sending it to the wire. This property can also can be set on the broker-side (so the messages are sent uncompressed by the producer, and compressed by the broker).
It depends on your network capacity to decide wether you prefer a slower producer (as the compression will slow it) or bigger load on the wire. Note that setting a big compression level on big files may affect a lot your overall performance.
Anyway, I believe the big/small msg problem hurts a lot more to the consumer side; Sending messages to Kafka is easy and fast (the default behaviour is async, so the producer won't be too busy). But on the consumer side, you'll have to look the way you are processing the messages:
One Consumer-Worker
Here you couple consuming with processing. This is the simplest way: the consumer sets its own thread, reads a kafka msg and process it. Then continues the loop.
One Consumer - Many workers
Here you decouple consuming and processing. In most cases, reading from kafka will be faster than the time you need to process the message. It is just physics. In this approach, one consumer feeds many separate worker threads that share the processing load.
More info about this here, just above the Constructors area.
Why do I explain this? Well, if your messages are too big, and you choose the first option, your consumer may not call poll() within the timeout interval, so it will rebalance continuosly. If your messages are big (and take some time to be processed), better choose to implement the second option, as the consumer will continue its own way, calling poll() without falling in rebalances.
If the messages are too big and too many, you may have to start thinking about different structures than can buffer the messages into your memory. Pools, deques, queues, for example, are different options to acomplish this.
You may also increase the poll timeout interval. This may hide you about dead consumers, so I don't really recommend it.
So my answer would be: it depends, basicallty on: your network capacity, your required latency, your processing capacity. If you are able to process big messages equally fast as smaller ones, then I wouldn't care much.
Maybe if you need to filter and reprocess older messages I'd recommend partitioning the topics and sending smaller messages, but it's only a use-case.
I am looking at using masstransit and have a need for selectively sending messages to consumers at the end if unreliable and slow network links (they are in the same WAN but use a slow and expensive cellular link).
I am expecting a fanout of 1 to 200 where the sites with lowest volume of messages and least reliable / most expensive links need to ignore the potentially high amount of message traffic othe consumers will see
I have looked at using the Selective consumer interface but this seems to imply that the message is always sent to all consumers, and then discarded if it doesn't match the predicate. This overhead is not acceptable.
Without using endpoint factory and manually managing uri end points to do a Send(), is there a nice way to do thus using subscriptions?
Simple answer: nope.
You do have a few options though. Is it just routing based upon load/processing? You could use competing consumers to do load balancing. All the endpoints read off the same queue (but they must be the same consumers on every process reading from the queue) and just pick up the next one. If you're slow, you just pick off fewer messages. (You can only use competing consumers with RabbitMQ).
For MSMQ there's a distributor that was built for load balancing. You could look at rebuilding that on top of RabbitMQ that if that's your transport. It's not super complicated, but would take some effort to do.
Other than that, I think you're likely down to writing something from scratch. It's not really pub/sub any more. So it falls outside MT's wheelhouse.
For a data-mining algorithm I am currently developing using Akka, I was wondering if Akka implements performance optimizations of the messages that are sent.
For instance, if I have an Actor that emits a very large number of messages to the same other Actor, is it good to encapsulate a set of messages into another large message? Or does Akka have some sort of buffer itself so that not one message but many messages are transfered over the network at once?
I am asking this question because the algorithm is supposed to be executed remotely on a cluster where transfer performance is important and I currently have no option to just do benchmarks myself.
For messages passed in Akka on the same machine, I don't think it matters a lot whether you use small message or an aggregation of messages as single message. The additional overhead of many calls versus having to loop while processing the aggregation is minimal I think.
I would prefer using small messages because it keeps the system simpler.
However, when sending messages over the network Akka is using HTTP and so there is the additional HTTP overhead costs for setting up a connection etc. Therefore you might choose here to aggregate some messages into a single message.
However, this also depends on your use case. Buffering implies waiting for more until there are enough (or a timeout occured). If you cannot wait, e.g. because you need fast responses, then you still need to send each message over individually.
I don't think there is a standard Akka actor available which does some aggregation of messages. Maybe a special kind of routing could be applied which does the buffering.
Or you might have a look at Akka Streams. That does support buffering of messages.
I'm trying to work out in my head the best way to structure a Cocoa app that's essentially a concurrent download manager. There's a server the app talks to, the user makes a big list of things to pull down, and the app processes that list. (It's not using HTTP or FTP, so I can't use the URL-loading system; I'll be talking across socket connections.)
This is basically the classic producer-consumer pattern. The trick is that the number of consumers is fixed, and they're persistent. The server sets a strict limit on the number of simultaneous connections that can be open (though usually at least two), and opening new connections is expensive, so in an ideal world, the same N connections are open for the lifetime of the app.
One way to approach this might be to create N threads, each of which would "own" a connection, and wait on the request queue, blocking if it's empty. Since the number of connections will never be huge, this is not unreasonable in terms of actual system overhead. But conceptually, it seems like Cocoa must offer a more elegant solution.
It seems like I could use an NSOperationQueue, and call setMaxConcurrentOperationCount: with the number of connections. Then I just toss the download requests into that queue. But I'm not sure, in that case, how to manage the connections themselves. (Just put them on a stack, and rely on the queue to ensure I don't over/under-run? Throw in a dispatch semaphore along with the stack?)
Now that we're in the brave new world of Grand Central Dispatch, does that open up any other ways of tackling this? At first blush, it doesn't seem like it, since GCD's flagship ability to dynamically scale concurrency (and mentioned in Apple's recommendations on Changing Producer-Consumer Implementations) doesn't actually help me. But I've just scratched the surface of reading about it.
EDIT:
In case it matters: yes, I am planning on using the asynchronous/non-blocking socket APIs to do the actual communication with the server. So the I/O itself does not have to be on its own thread(s). I'm just concerned with the mechanics of queuing up the work, and (safely) doling it out to the connections, as they become available.
If you're using CFSocket's non-blocking calls for I/O, I agree, that should all happen on the main thread, letting the OS handle the concurrency issues, since you're just copying data and not really doing any computation.
Beyond that, it sounds like the only other work your app needs to do is maintain a queue of items to be downloaded. When any one of the transfers is complete, the CFSocket call back can initiate the transfer of the next item on the queue. (If the queue is empty, decrement your connection count, and if something is added to an empty queue, start a new transfer.) I don't see why you need multiple threads for that.
Maybe you've left out something important, but based on your description the app is I/O bound, not CPU bound, so all of the concurrency stuff is just going to make more complicated code with minimal impact on performance.
Do it all on the main thread.
For posterity's sake, after some discussion elsewhere, the solution I think I'd adopt for this is basically:
Have a queue of pending download operations, initially empty.
Have a set containing all open connections, initially empty.
Have a mutable array (queue, really) of idle open connections, initially empty.
When the user adds a download request:
If the array of idle connections is not empty, remove one and assign the download to it.
If there are no idle connections, but the number of total connections has not reached its limit, open a new connection, add it to the set, and assign the download to it.
Otherwise, enqueue the download for later.
When a download completes: if there are queued requests, dequeue one
and give it to the connection; otherwise, add the connection to the idle list.
All of that work would take place on the main thread. The work of decoding the results of each download would be offloaded to GCD, so it can handle throttling the concurrency, and it doesn't clog the main thread.
Opening a new connection might take a while, so the process of creating a new one might be a tad more complicated in actual practice (say, enqueue the download, initiate the connection process, and then dequeue it when the connection is fully established). But I still think my perception of the possibility of race conditions was overstated.
I'm going through a bit of a re-think of large-scale multiplayer games in the age of Facebook applications and cloud computing.
Suppose I were to build something on top of existing open protocols, and I want to serve 1,000,000 simultaneous players, just to scope the problem.
Suppose each player has an incoming message queue (for chat and whatnot), and on average one more incoming message queue (guilds, zones, instances, auction, ...) so we have 2,000,000 queues. A player will listen to 1-10 queues at a time. Each queue will have on average maybe 1 message per second, but certain queues will have much higher rate and higher number of listeners (say, a "entity location" queue for a level instance). Let's assume no more than 100 milliseconds of system queuing latency, which is OK for mildly action-oriented games (but not games like Quake or Unreal Tournament).
From other systems, I know that serving 10,000 users on a single 1U or blade box is a reasonable expectation (assuming there's nothing else expensive going on, like physics simulation or whatnot).
So, with a crossbar cluster system, where clients connect to connection gateways, which in turn connect to message queue servers, we'd get 10,000 users per gateway with 100 gateway machines, and 20,000 message queues per queue server with 100 queue machines. Again, just for general scoping. The number of connections on each MQ machine would be tiny: about 100, to talk to each of the gateways. The number of connections on the gateways would be alot higher: 10,100 for the clients + connections to all the queue servers. (On top of this, add some connections for game world simulation servers or whatnot, but I'm trying to keep that separate for now)
If I didn't want to build this from scratch, I'd have to use some messaging and/or queuing infrastructure that exists. The two open protocols I can find are AMQP and XMPP. The intended use of XMPP is a little more like what this game system would need, but the overhead is quite noticeable (XML, plus the verbose presence data, plus various other channels that have to be built on top). The actual data model of AMQP is closer to what I describe above, but all the users seem to be large, enterprise-type corporations, and the workloads seem to be workflow related, not real-time game update related.
Does anyone have any daytime experience with these technologies, or implementations thereof, that you can share?
#MSalters
Re 'message queue':
RabbitMQ's default operation is exactly what you describe: transient pubsub. But with TCP instead of UDP.
If you want guaranteed eventual delivery and other persistence and recovery features, then you CAN have that too - it's an option. That's the whole point of RabbitMQ and AMQP -- you can have lots of behaviours with just one message delivery system.
The model you describe is the DEFAULT behaviour, which is transient, "fire and forget", and routing messages to wherever the recipients are. People use RabbitMQ to do multicast discovery on EC2 for just that reason. You can get UDP type behaviours over unicast TCP pubsub. Neat, huh?
Re UDP:
I am not sure if UDP would be useful here. If you turn off Nagling then RabbitMQ single message roundtrip latency (client-broker-client) has been measured at 250-300 microseconds. See here for a comparison with Windows latency (which was a bit higher) http://old.nabble.com/High%28er%29-latency-with-1.5.1--p21663105.html
I cannot think of many multiplayer games that need roundtrip latency lower than 300 microseconds. You could get below 300us with TCP. TCP windowing is more expensive than raw UDP, but if you use UDP to go faster, and add a custom loss-recovery or seqno/ack/resend manager then that may slow you down again. It all depends on your use case. If you really really really need to use UDP and lazy acks and so on, then you could strip out RabbitMQ's TCP and probably pull that off.
I hope this helps clarify why I recommended RabbitMQ for Jon's use case.
I am building such a system now, actually.
I have done a fair amount of evaluation of several MQs, including RabbitMQ, Qpid, and ZeroMQ. The latency and throughput of any of those are more than adequate for this type of application. What is not good, however, is queue creation time in the midst of half a million queues or more. Qpid in particular degrades quite severely after a few thousand queues. To circumvent that problem, you will typically have to create your own routing mechanisms (smaller number of total queues, and consumers on those queues are getting messages that they don't have an interest in).
My current system will probably use ZeroMQ, but in a fairly limited way, inside the cluster. Connections from clients are handled with a custom sim. daemon that I built using libev and is entirely single-threaded (and is showing very good scaling -- it should be able to handle 50,000 connections on one box without any problems -- our sim. tick rate is quite low though, and there are no physics).
XML (and therefore XMPP) is very much not suited to this, as you'll peg the CPU processing XML long before you become bound on I/O, which isn't what you want. We're using Google Protocol Buffers, at the moment, and those seem well suited to our particular needs. We're also using TCP for the client connections. I have had experience using both UDP and TCP for this in the past, and as pointed out by others, UDP does have some advantage, but it's slightly more difficult to work with.
Hopefully when we're a little closer to launch, I'll be able to share more details.
Jon, this sounds like an ideal use case for AMQP and RabbitMQ.
I am not sure why you say that AMQP users are all large enterprise-type corporations. More than half of our customers are in the 'web' space ranging from huge to tiny companies. Lots of games, betting systems, chat systems, twittery type systems, and cloud computing infras have been built out of RabbitMQ. There are even mobile phone applications. Workflows are just one of many use cases.
We try to keep track of what is going on here:
http://www.rabbitmq.com/how.html (make sure you click through to the lists of use cases on del.icio.us too!)
Please do take a look. We are here to help. Feel free to email us at info#rabbitmq.com or hit me on twitter (#monadic).
My experience was with a non-open alternative, BizTalk. The most painful lesson we learnt is that these complex systems are NOT fast. And as you figured from the hardware requirements, that translates directly into significant costs.
For that reason, don't even go near XML for the core interfaces. Your server cluster will be parsing 2 million messages per second. That could easily be 2-20 GB/sec of XML! However, most messages will be for a few queues, while most queues are in fact low-traffic.
Therefore, design your architecture so that it's easy to start with COTS queue servers and then move each queue (type) to a custom queue server when a bottleneck is identified.
Also, for similar reasons, don't assume that a message queue architecture is the best for all comminication needs your application has. Take your "entity location in an instance" example. This is a classic case where you don't want guaranteed message delivery. The reason that you need to share this information is because it changes all the time. So, if a message is lost, you don't want to spend time recovering it. You'd only send the old locatiom of the affected entity. Instead, you'd want to send the current location of that entity. Technology-wise this means you want UDP, not TCP and a custom loss-recovery mechanism.
FWIW, for cases where intermediate results are not important (like positioning info) Qpid has a "last-value queue" that can deliver only the most recent value to a subscriber.