I need to submit several thousand jobs to our cluster. Each job needs around six hours to complete. This will take around a week if I would use all available resources. Theoretically I could do that but the I would block all other users for a week. So this is not an option.
I have two ideas that could possibly solve the problem:
Create an array job and limit the maximum number of running jobs. I don't like this option because quite often (over night, weekends, etc.) no one uses the cluster and my jobs can not use these unused resources.
Submit all jobs at once but somehow set the priority of each job really low. Ideally anyone could still use the cluster because when they submit jobs they will start sooner than mine. I do not know if this is possible in slurm and if I would have the permission to do that.
Is there a slurm mechanism I am missing? Is it possible to set priorities of a slurm job as described above and would I have permission to do that?
Generaly this is the cluster admin problem. They should have configured the cluster in a way that prioritize short and small jobs over long and large ones and/or prevent large jobs from running on some nodes.
However you can also manually reduce the priority of your job as a non admin with the nice factor option (higher -> less priority):
sbatch --nice=POSITIVE_NUMBER script.sh
Related
I'm loading parquet data into a dataframe via
spark.read.parquet('hdfs:///path/goes/here/...')
There are around 50k files in that path due to parquet partitioning. When I run that command, spark spawns off dozens of small jobs that as a whole take several minutes to complete. Here's what the jobs look like in the spark UI:
As you can see, although each job has ~2100 tasks, they execute quickly, in about 2 seconds. Starting so many 'mini jobs' is inefficient and leads this file listing step to take about 10 minutes (where the clusters resources are mostly idle, and the cluster is mostly dealing with straggling tasks or the overhead of managing jobs/tasks).
How can I consolidate these tasks into fewer jobs, each with more tasks?
Bonus points for a solution that also works in pyspark.
I'm running spark 2.2.1 via pyspark on hadoop 2.8.3.
I believe you encountered a bug for which a former colleague of mine has filed a ticket and opened a pull request. You can check it out here. If it fits your issue, your best shot is probably voting the issue up and making some noise on the mailing list about it.
What you might want to do is tweaking the spark.sql.sources.parallelPartitionDiscovery.threshold and spark.sql.sources.parallelPartitionDiscovery.parallelism configuration parameters (with the former being cited in the linked ticket) in a way that suits your job.
You can have a look here and here to see how the configuration key is used. I'll share the related snippets here for completeness.
spark.sql.sources.parallelPartitionDiscovery.threshold
// Short-circuits parallel listing when serial listing is likely to be faster.
if (paths.size <= sparkSession.sessionState.conf.parallelPartitionDiscoveryThreshold) {
return paths.map { path =>
(path, listLeafFiles(path, hadoopConf, filter, Some(sparkSession)))
}
}
spark.sql.sources.parallelPartitionDiscovery.parallelism
// Set the number of parallelism to prevent following file listing from generating many tasks
// in case of large #defaultParallelism.
val numParallelism = Math.min(paths.size, parallelPartitionDiscoveryParallelism)
The default value for this configuration are 32 for the threshold and 10000 for the parallelism (related code here).
In your case, I'd say that probably what you want to do is setting the threshold so that the process is run without spawning parallel jobs.
Note
The linked sources are from the latest available tagged release at the time of writing, 2.3.0.
Against an object store, even the listing and calls to getFileStatus are pretty expensive, and as this is done during partitioning, can extend the job a lot.
Play with mapreduce.input.fileinputformat.list-status.num-threads to see if adding more threads speeds things up, say a value of 20-30
I am currently developing an application to wrangle a huge amount of data using Spark. The data is a mixture of Apache (and other) log files as well as csv and json files. The directory structure of my Google bucket will look something like this:
root_dir
web_logs
\input (subdirectory)
\output (subdirectory)
network_logs (same subdirectories as web_logs)
system_logs (same subdirectories as web_logs)
The directory structure under the \input directories is arbitrary. Spark jobs pick up all of their data from the \input directory and place it in the \output directory. There is an arbitrary number of *_logs directories.
My current plan is to split the entire wrangling task into about 2000 jobs and use the cloud dataproc api to spin up a cluster, do the job, and close down. Another option would be to create a smaller number of very large clusters and just send jobs to the larger clusters instead.
The first approach is being considered because each individual job is taking about an hour to complete. Simply waiting for one job to finish before starting the other will take too much time.
My questions are: 1) besides the cluster startup costs, are there any downside to taking the first approach? and 2) is there a better alternative?
Thanks so much in advance!
Besides startup overhead, the main other consideration when using single-use clusters per job is that some jobs might be more prone to "stragglers" where data skew leads to a small number of tasks taking much longer than other tasks, so that the cluster isn't efficiently utilized near the end of the job. In some cases this can be mitigated by explicitly downscaling, combined with the help of graceful decommissioning, but if a job is shaped such that many "map" partitions produce shuffle output across all the nodes but there are "reduce" stragglers, then you can't safely downscale nodes that are still responsible for serving shuffle data.
That said, in many cases, simply tuning the size/number of partitions to occur in several "waves" (i.e. if you have 100 cores working, carving the work into something like 1000 to 10,000 partitions) helps mitigate the straggler problem even in the presence of data skew, and the downside is on par with startup overhead.
Despite the overhead of startup and stragglers, though, usually the pros of using new ephemeral clusters per-job vastly outweigh the cons; maintaining perfect utilization of a large shared cluster isn't easy either, and the benefits of using ephemeral clusters includes vastly improved agility and scalability, letting you optionally adopt new software versions, switch regions, switch machine types, incorporate brand-new hardware features (like GPUs) if they become needed, etc. Here's a blog post by Thumbtack discussing the benefits of such "job-scoped clusters" on Dataproc.
A slightly different architecture if your jobs are very short (i.e. if each one only runs a couple minutes and thus amplify the downside of startup overhead) or the straggler problem is unsolveable, is to use "pools" of clusters. This blog post touches on using "labels" to easily maintain pools of larger clusters where you still teardown/create clusters regularly to ensure agility of version updates, adopting new hardware, etc.
You might want to explore my solution for Autoscaling Google Dataproc Clusters
The source code can be found here
I am working with NGS data and the newest test files are massive.
Normally our pipeline is using just one node and the output from different tools is its ./scratch folder.
To use just one node is not possible with the current massive data set. That's why I would like to use at least 2 nodes to solve the issues such as speed, not all jobs are submitted, etc.
Using multiple nodes or even multiple partitions is easy - i know how which parameter to use for that step.
So my issue is not about missing parameters, but the logic behind slurm to solve the following issue about I/O:
Lets say I have tool-A. Tool-A is running with 700 jobs on two nodes (340 jobs on node1 and 360 jobs on node2) - the ouput is saved on ./scratch on each node separately.
Tool-B is using the results from tool-A - which are on two different nodes.
What is the best approach to fix that?
- Is there a parameter which tells slurm which jobs belongs together and where to find the input for tool-B?
- would it be smarter to change the output on /scratch to a local-folder?
- or would it be better to merge the output from tool-A from both nodes to one node?
- any other ideas?
I hope I made my issue "simply" to understand... Please apologize if that is not the case!
My naive suggestion would be why not share a scratch nfs volume across all nodes ? This way all ouput datas of ToolA would be acessible for ToolB whatever the node. It migth not be the best solution for read/write speed, but to my mind it would be the easiest for your situation.
A more sofware solution (not to hard to develop) can be to implement a database that track where the files have been generated.
I hope it help !
... just for those coming across this via search engines: if you cannot use any kind of shared filesystem (NFS, GPFS, Lustre, Ceph) and you don't have only massive data sets, you could use "staging", meaning data transfer before and after your job really runs.
Though this is termed "cast"ing in the Slurm universe, it generally means you define
files to be copied to all nodes assigned to your job BEFORE the job starts
files to be copied from nodes assigned to your job AFTER the job completes.
This can be a way to get everything needed back and forth from/to your job's nodes even without a shared file system.
Check the man page of "sbcast" and amend your sbatch job scripts accordingly.
I am running a spark direct stream from kafka where I need to run many concurrent jobs in order to process all the data in time. In spark you can set spark.streaming.concurrentJobs to a number of concurrent jobs you want to run.
What I want to know is a logical way to determine how many concurrent jobs I can run within my given environment. For privacy issues at my company, I cannot tell you the specs that I have, but what I would want to know is which specs are relevant in determining a limit and why?
Of course the alternative is that I could keep increasing it and testing, then adjusting based on results but I would like a more logical approach and I want to actually understand what determines that limit and why.
To test different numbers of concurrent jobs and see the overall execution time is the most reliable method. However, I suppose the best number roughly equals to the value of Runtime.getRuntime().availableProcessors();
So my advice is to start with that number of available processors, then increase and decrease it by 1,2, and 3. Then make a chart (execution time against the number of jobs) and you'll see the optimal number of jobs.
Lets say my job performs several spark actions, where the first few are not using multiple cores for a single task so I would like each instance to perform (executor.cores) tasks in parallel (spark.task.cpus=1).
Then suppose I have another action which can be parallelized - I'm desiring a feature where I could increase spark.task.cpus (say to use more cores on the executor), and perform fewer tasks simultaneously on each instance.
My workaround right now is to save data, start a new sparkContext with new settings, and reload the data.
The use case: my later actions may be unavoidable skewed and I may want to apply more than one core per task to avoid bottlenecking on such large tasks, but I don't want this to impact the earlier actions which can benefit from using 1 core per task.
From looking around my guess is that I can't do this currently, so I'm mainly wondering if there is a a significant limitation for not allowing this. Alternatively, suggestions for how I could trick spark into achieving something similar.
Note: Currently using 1.6.2 but willing to hear other options for Spark2+