I want to develop a task scheduler using MPI where there is a single master processor and there are worker/client processors. Each worker has all the data it needs to compute, but gets the index to work on from the master. After the computation the worker returns some data to the master. The problem is that some processes will be fast and some will be slow.
If I run a loop so that at each iteration the master sends and receives (blocking/non-blocking) data then it can't proceed to next step till it has received data from the current worker from the previous index assigned to it. The bottom line is if a worker takes too long to compute then it becomes the limiting factor and the master can't move on to assign an index to the next worker even if non-blocking techniques are used. Is it possible to skip assigning to a worker and move on to next.
I'm beginning to think that MPI might not be the paradigm to do this. Would python be a nice platform to do task scheduling?
This is absolutely possible using MPI_Irecv() and MPI_Test(). All the master process needs to do is post a non-blocking receive for each worker process, then in a loop test each one for incoming data. If a process is done, send it a new index, post a new non-blocking receive for it, and continue.
One MPI_IRecv for each process is one solution. This has the downside of needing to cancel unmatched MPI_IRecv when the work is complete.
MPI_ANY_SOURCE is an alternate path. This will allow the manager process to have a single MPI_IRecv outstanding at any given time, and the "next" process to MPI_Send will be matched with MPI_ANY_SOURCE. This has the downside of several ranks blocking in MPI_Send when there is no additional work to be done. Some kind of "nothing more to do" signal needs to be worked out, so the ranks can do a clean exit.
Related
This is purely for a non-eager pytest mode of operation. I want to know when celery has "caught" up with all the outstanding work. Is there any way to find that information? My testing config has a celery_session_app and a single celery_session_worker in it's own thread.
Check the number of entries in the Rabbit queue. This has problems because of pre-fetch. I can set prefetch to 1 and maybe solve it that way but I worry about race conditions. (I'm testing chords and some celery tasks queue other celery tasks)
Add a task to the "end" of the list and then .wait() on it to finish. This has problems for tasks that queue other tasks because the queue is being extended in the other thread so I can be at the end of the list when queued, but that quickly moves forward as tasks are queued behind it. I can work around this using .apply_async(countdown=3) but this is pretty much the definition of a race condition and I might need countdown=4 or I might need nothing and that is some number of seconds wasted on a test regardless.
Use signals (somehow). But what I really need is a worker_is_bored which does not exist and suffers from the same kind of race conditions mentioned above. Tasks queueing tasks could make it flash "bored" and right back to "busy".
time.sleep(N) but what should N be. (i'm running pytest -n 10 so how busy the machine is during tests, is non-trivial). And this wastes time like countdown= above.
I have a job processing system where each job contains thousands of individual tasks that require different strategies to complete. The individual tasks make up the whole job. If all tasks have been completed, the job is marked as successfully completed and other steps are taken, if any of the tasks fail, the job must be marked as failed and other steps are taken, if the job times out the job must be marked as failed and other steps are taken.
Once all of the results for a job have been received, the next job can be fetched. The next job shouldn't be fetched while a job is currently being processed.
Here is the what the flow looks like:
The Job Polling Verticle publishes a job to the event bus, and the Job Processing Verticle publishes each task to the event bus. When the job strategy completes, it publishes the task result to the event bus.
The issue is that I don't know the right way to determine when all tasks have been completed in this model. All verticles are stateless, The Job Processing Verticle doesn't await any futures, and even if the Job Results Verticle was stateful, it doesn't know how many results it should expect.
The only way I can think to do this would be to have a global stateful object. But I don't think this is good design.
Additionally, I need to know when a Job has timed out. That is, it's run longer than it should and I need to consider it's failed, log it, and move on.
I could do this with the global state, but again I don't think that's the right solution.
Does this verticle pattern make sense for what I'm trying to do?
First, let me try to address your questions. Then I'll try to explain what problems this design has.
The issue is that I don't know the right way to determine when all tasks have been completed in this model. All verticles are stateless, The Job Processing Verticle doesn't await any futures, and even if the Job Results Verticle was stateful, it doesn't know how many results it should expect.
The solution could be reference counting verticle. Each worker should emit a start message on event bus with jobId when it starts, and end message with jobId when it completes. Even if you have fan-out (those are the cases that you don't know how many workers there are), counting verticle will know that. In your diagram, "Job Post Processing Verticle" is a good candidate for this. It can maintain a counter, and only when it reaches zero, it should start the next job. That also helps avoiding actually sharing some memory reference.
Additionally, I need to know when a Job has timed out. That is, it's run longer than it should and I need to consider it's failed, log it, and move on.
In the same verticle you can start a timer every time you get a new start message. If you get end message, cancel the timer. Otherwise, cancel current job and start again.
Now, this solution will work, but the design has two main flaws. One is the fact that you maintain all your flow in memory, it seems. If your application crashes, all progress is lost, and it's not clear how you record it. Maybe polling Jobs table in DB would actually be better, since your job execution is sequential anyway.
Second point is the fact that all those timeouts and reference counting is homemade implementation of structured concurrency. Maybe you should take a look at something like Kotlin coroutines for that, at it will handle many of your problems for you.
I've been using supervisord for a while -- outstanding tool. The one use case I haven't been able to figure out is, how to configure jobs to be restarted until a condition is met, then stop restarting.
Example: let's say you have a bunch of work to do, like scaling thousands of images, or servicing millions of requests on a queue. A useful pattern would be to run many workers in parallel to work on that backlog. You could set up a supervisord job that ensures 100 workers are running, and if any of them crash, supervisord will spin up replacements so the pool of workers won't shrink.
That's great until the work is done. Maybe when the backlog is gone, the number of workers should scale down to 1 or 0. Supervisord will keep spinning up the total to be 100 processes, even if each new process checks to see if there's work to be done, sees none, and shuts down very quickly.
Is there a way for a process instance or process family to communicate with supervisord to say, the autoretsart behavior is no longer needed? Better yet, is there a way to scale the number of worker processes up and down based on some condition (like number of files in a directory or ??).
I know it can be done by updating the supervisord.conf file and running supervisorctl reload, but I'd prefer something that's more declarative and self-managing if such a thing exists.
Is there a way for a process instance or process family to communicate with supervisord to say, the autoretsart behavior is no longer needed?
You can wind down an activity by making sure your processes exit with different exitcode(s) when there is no work and making those the expected exitcodes with autorestart=unexpected in the configuration.
Better yet, is there a way to scale the number of worker processes up and down based on some condition (like number of files in a directory or ??).
The trouble is that the automatic state transitions don't allow for getting processes running again from an expected EXITED state. AFAIK the only way to do this is with the XML-RPC API's startProcess, so you would need to write or find an appropriate event listener that watches for your start condition and then uses the API.
An alternate design is to wrap your worker process in an event handler watching PROCESS COMMUNICATION Events and have one normal subprocess communicating new tasks to a pool of event listeners. But that model doesn't currently eliminate a pool of waiting processes when there is no work, it just organizes the control task in a way that may make it easier to separate out task related logic and resource usage.
So, i built this small example of a ZeroMQ pipeline architecture because i'll end up having to do something similar very soon and i'm trying to grasp the pipeline concept the right way.
https://gist.github.com/2765708
Right now, this is completely asynchronous. The controller dispatches a batch of tasks to various workers, which in their turn, send a message to the sink. The controller and sink are fixed parts of my architecture, while workers are dynamic. That's perfect.
However, i would like to know when the workers have finished working on all their tasks. In that example, i do know the amount of messages, but that won't be true on real-life situations. I might have 100 messages or 10,000. So, how can the sink or the controller know when the workers have finished working on their tasks? I have to perform some actions that depend on the conclusion of the jobs sent to workers.
I wanted to expand on #bjlaub's answer. It started as a comment but I was typing too much. I agree with the concept of acknowledgment, but believe it can originate in multiple places.
There are multiple approaches to this communication and it all depends on the behavior you are after in the system.
First, you can either send out messages from the workers as they finish each task, or from the sink as it receives each task. Right now I am not addressing the type of socket, only the act of communicating. I believe it is much more efficient to send it from the sink as you would only need one connection back to the controller instead of one for each worker. The sink does not need to know how many total tasks there are. Only that it is firing off a message after each result it receives. The controller can determine how many to expect since it was the submission point and new when it had exhausted its submission (the count).
Now regardless of whether you have the message sent from the worker or the sink, you can use different socket types. If you want the controller to completely block until all work is done, then you can have it be a push/pull until it receives X messages (message content can be anything. Its just a trigger).
This may be limiting if the controller wants to be able to do other work while these tasks are happening. If so, you could maybe use pub/sub, and let the controller subscribe to being notified as tasks complete, and asynchronously maintain a count until the total has been satisfied.
And finally, maybe you have the situation where you want the controller to ask the sink for a status when you deem fit. You can have a req/rep pattern for the controller to ask the sink how many requests it has received on demand.
I'm sure one of these patterns will fit your specific needs.
One idea (disclaimer: I have very little experience w/ 0MQ!):
Setup an "acknowledgment" pipeline in the reverse direction. Since the controller presumably knows how many tasks it has dispatched to the workers (e.g. the number of times it called send), it can use a PULL socket to receive a small message (an integer for example) from each worker indicating the completion of the task. The worker process dispatches its completed result to the sink, and at the same time sends the acknowledgement back to the controller. Once the controller collects the right number of acknowledgements, it can do whatever post-processing is necessary before farming out the next set of work.
You could also push this downstream to the sink, but you would need to notify the sink of the total number of work units to expect before farming them out to the workers.
I'm working on a system that uses several hundreds of workers in parallel (physical devices evaluating small tasks). Some workers are faster than others so I was wondering what the easiest way to load balance tasks on them without a priori knowledge of their speed.
I was thinking about keeping track of the number of tasks a worker is currently working on with a simple counter and then sorting the list to get the worker with the lowest active task count. This way slow workers would get some tasks but not slow down the whole system. The reason I'm asking is that the current round-robin method is causing hold up with some really slow workers (100 times slower than others) that keep accumulating tasks and blocking new tasks.
It should be a simple matter of sorting the list according to the current number of active tasks, but since I would be sorting the list several times a second (average work time per task is below 25ms) I fear that this might be a major bottleneck. So is there a simple version of getting the worker with the lowest task count without having to sort over and over again.
EDIT: The tasks are pushed to the workers via an open TCP connection. Since the dependencies between the tasks are rather complex (exclusive resource usage) let's say that all tasks are assigned to start with. As soon as a task returns from the worker all tasks that are no longer blocked are queued, and a new task is pushed to the worker. The work queue will never be empty.
How about this system:
Worker reaches the end of its task queue
Worker requests more tasks from load balancer
Load balancer assigns N tasks (where N is probably more than 1, perhaps 20 - 50 if these tasks are very small).
In this system, since you are assigning new tasks when the workers are actually done, you don't have to guess at how long the remaining tasks will take.
I think that you need to provide more information about the system:
How do you get a task to a worker? Does the worker request it or does it get pushed?
How do you know if a worker is out of work, or even how much work is it doing?
How are the physical devices modeled?
What you want to do is avoid tracking anything and find a more passive way to distribute the work.