In my application, I sometimes get into a state where work isn't getting done by my workers, but the CPU and disk are basically sitting idle. I'd like to be able to regular log (e.g. to statsd or similar) the number of active worker threads and maximum number of active worker threads for each scheduler. Then if we have problems, we can check the logs and cross-reference to determine if the problems coincided with our thread pools being totally full.
I can't seem to find any methods to determine, at the current moment in time, the total thread pool size and number of running threads in each pool. Where should I look?
Related
I am evaluating Batch for a project and, while it seems like it will do what I'm looking for, I'm not sure if what I am assuming is really correct.
I have what is basically a job runner from a queue. The current solution works but when the pool of nodes scales down, it just blindly kills off machines. I am looking for something that, when scaling down, will allow currently-running jobs to complete and then remove the node(s) from the pool. I also want to preemptively increase the pool size if a spike is likely to occur (and not have those nodes shut down). I can adjust the pool size externally if that makes sense (seems like the best option so far).
My current idea is to have one pool with one job & task per node, and that task listens to a queue in a loop for messages and processes them. After an iteration count and/or time limit, it shuts down, removing that node from the pool. If the pool size didn't change, I would like to replace that node with a new one. If the pool was shrunk, it should just go away. If the pool size increases, new nodes should run and start up the task.
I'm not planning on running something that continually add pools, or nodes to the pool, or tasks to a job, though I will probably have something that sets the pool size periodically based on queue length or something similar. What I would rather not do is something like "there are 10 things in the queue, add a pool with x nodes, then delete it".
Is this possible or are my expectations incorrect? So far, from reading the docs, it seems like it should be doable, and I have a simple task working, but I'm not sure about the scaling mechanics or exactly how to structure the tasks/jobs/pools.
Here's one possible way to lean into the strengths of Azure Batch and achieve what you've described.
Create your job with a JobManagerTask that monitors your queue for incoming work and adds a new Batch Task for each item of your workload. Each task will process a single piece of work, then exit.
The Batch Scheduler will then take care of allocating tasks to compute nodes. It can also take care of retrying tasks that fail, and so on.
Configure your pool with an AutoScale formula to dynamically resize your pool to meet your load. Your formula can specify taskcompletion to ensure tasks get to complete before any one compute node is removed.
If your workload peaks are predictable (say, 9am every day) your AutoScale expression could scale up your pool in anticipation. If those spikes are not predicable, your external monitoring (or your JobManager) can change the AutoScale expression at any time to suit.
If appropriate, your job manager can terminate once all the required tasks have been added; set onAllTasksComplete to terminatejob, ensuring your job is completed after all your tasks have finished.
A single pool can process tasks from multiple jobs, so if you have multiple concurrent workloads, they could share the same pool. You can give jobs different values for priority if you want certain jobs to be processed first.
In an Operating System, threads are typically handled in user mode or kernel mode. What are some of the advantages and disadvantages of each?
User-mode threads are scheduled in user mode by something in the process, and the process itself is the only thing handled by the kernel scheduler.
That means your process gets a certain amount of grunt from the CPU and you have to share it amongst all your user mode threads.
Simple case, you have two processes, one with a single thread and one with a hundred threads.
With a simplistic kernel scheduling policy, the thread in the single-thread process gets 50% of the CPU and each thread in the hundred-thread process gets 0.5% each.
With kernel mode threads, the kernel itself manages your threads and schedules them independently. Using the same simplistic scheduler, each thread would get just a touch under 1% of the CPU grunt (101 threads to share the 100% of CPU).
In an Operating System, threads are typically handled in user mode or kernel mode.
Typically threads are handled in kernel mode.
What are some of the advantages and disadvantages of each?
In theory, the advantage of handling threads in user mode is that it avoids the cost of switching to/from kernel when a thread needs to wait for something (which can be relatively expensive as it involves privilege level switches). In practice this "advantage" often doesn't happen because the thread has to switch to kernel anyway, to ask kernel to do whatever the thread would wait for (e.g. switching to kernel to ask it to read data from a file and then returning to user-space to block/wait instead of blocking/waiting in the kernel while you're already in the kernel). Mostly; it only helps if the kernel isn't involved at all, which only really happens when user-space threads communicate with or share locks with other threads in the same process.
The advantage of handling threads in kernel is that the kernel can support thread priorities properly. For example, if you have two processes that both have a very high priority thread and a very low priority thread; then kernel can make sure CPU time is given to the high priority thread/s when possible (including pre-empting low priority threads when a high priority thread unblocks) because it knows about all threads; but user-space can't do this - one process doesn't know about threads belonging to a different process, so user threading will get it wrong and ruin performance (one process giving CPU time to its own very low priority thread while a very high priority thread belonging to a different process needs the CPU and doesn't get it).
The other advantage of handling threads in the kernel is that (especially for systems with multiple CPUs) the kernel has access to better information and can make smarter scheduling decisions. This includes balancing the load (from any number of processes) across all CPUs while taking into account "CPU topology" (NUMA, SMT, etc; possibly including heterogeneous CPUs - e.g. "big.LITTLE" arrangements); and making trade-offs between thread priorities, CPU temperatures and power consumption (e.g. if one of the CPU's is getting too hot, reduce that CPU's clock speed to let it cool down and use it for low priority threads so that the performance of high priority threads isn't effected).
I am using a cluster of 12 virtual machines, each of which has 16 GB memory and 6 cores(except master node with only 2 cores). To each worker node, 12GB memory and 4 cores were assigned.
When I submit a spark application to yarn, I set the number of executors to 10(1 as master manager, 1 as application master), and to maximize the parallelism of my application, most of my RDDs have 40 partitions as same as the number of cores of all executors.
The following is the problem I encountered: in some random stages, some tasks need to be processed extremely longer than others, which results in poor parallelism. As we can see in the first picture, executor 9 executed its tasks over 30s while other tasks could be finished with 1s. Furthermore, the reason for much time consumed is also randomized, sometimes just because of computation, but sometimes scheduler delay, deserialization or shuffle read. As we can see, the reason for second picture is different from first picture.
I am guessing the reason for this occurs is once some task got assigned to a specific slot, there is not enough resources on the corresponding machine, so jvm was waiting for cpus. Is my guess correct? And how to set the configuration of my cluster to avoid this situation?
computing
scheduler delay & deserialization
To get a specific answer you need to share more about what you're doing but most likely the partitions you get in one or more of your stages are unbalanced - i.e. some are much bigger than others. The result is slowdown since these partitions are handled by a specific task. One way to solve it is to increase the number of partitions or change the partitioning logic
When a big task finishes shipping the data to other tasks would take longer as well so that's why other tasks may take long
we're having a setup with a large number of separate django installations on a single box. each of these have their own code base & linux user.
we're using celery for some asynchronous tasks.
each of the installations has its own setup for celery, i.e. its own celeryd & worker.
the amount of asynchronous tasks per installation is limited, and not time-critical.
when a worker starts it takes about 30mb of memory. when it has run for a while this amount may grow (presumably due to fragmentation).
the last bulletpoint has already been (somewhat) solved by settings --maxtasksperchild to a low number (say 10). This ensures a restart after 10 tasks, after which the memory at least goes back to 30MB.
However, each celeryd is still taking up a lot of memory, since the minimum amount of workers appears to be 1 as opposed to 0. I also imagine running python manage.py celery worker does not lead to the smallest-possible footprint for the celeryd, since the full stack is loaded even if the only thing that happens is checking for tasks.
In an ideal setup, I'd like to see the following: a process that has a very small memory footprint (100k or so) is looking at the queue for new tasks. when such a task arises, it spins up the (heavy) full django stack in a separate process. and when the worker is done, the heavy process is spun down.
Is such a setup configurable using (somewhat) standard celery? If not, what points of extension are there?
we're (currently) using celery 3.0.17 and the associated django-celery.
Just to make sure I understand - you have a lot of different django codebases, each with their own celery, and they take up too much memory when running on a single box simultaneously, all waiting for a celery job to come down the pipe? How many celery instances are we talking about here?
In my experience, you're using django celery in a very different way than it was designed for - all of your different django projects should be condensed to a few (or a single) project(s), composed of multiple applications. Then you set up a small number of queues to field celery tasks from the different apps - this way, you only have as many dormant celery threads taking up 30mb as you have queues, and a single queue can handle multiple tasks (from multiple apps if you want). The memory issue should go away.
To reiterate - you only need one celeryd, driving multiple workers. This way your bottleneck is job concurrency, not dormant memory needs.
Why do you need so many django installations? Please let me know if I'm missing something, or if you need clarification.
Could anyone point to me what is the overhead of running a matlabpool ?
I started a matlabpool :
matlabpool open 132procs 100
Starting matlabpool using the '132procs' configuration ... connected to 100 labs.
And followed cpu usage on the nodes as :
pdsh -A ps aux |grep dmlworker
When I launch the matlabpool, it starts with ~35% cpu usage on average and when the pool
is not being used it slowly (in 5-7 minutes) goes down to ~2% on average.
Is this normal ? What is the typical overhead ? Does that change if matlabpooljob is launched as a "batch" job ?
This is normal. ps aux reports the average CPU utilization since the process was started, not over a rolling window. This means that, although the workers initialize relatively quickly and then become idle, it will take longer for this to reflect in CPU%. This is different to the Linux top command, for example, which will reflect the utilization since the last screen update in %CPU.
As for typical overhead, this depends on a number of factors: clearly the number of workers, the rate and data size of jobs submitted (as well as in maintaining the worker processes, there is some overhead in marshalling input and output, which is not part of "useful computation"), whether the Matlab pool is local or attached to a job manager, and the Matlab version and O/S.
From experience, as a rough guide on a modern *nix server, I would think an idle worker should be not be consuming more than 20% of a single core (e.g. <~1% total CPU utilization on a 16-core box) after initilization, unless there is a configuration issue. I should not expect this to be influenced by what kind of jobs you are submitting (whether using "createJob" or "batch" or "parfor" for example): the workers and communication mechanisms underneath are essentially the same.