Pipeline Optimization » History » Revision 17
Revision 16 (Bryan Cosca, 04/19/2016 02:47 PM) → Revision 17/31 (Bryan Cosca, 04/20/2016 07:51 PM)
h1. Pipeline Optimization h2. Overview This wiki page is designed to help users make their Arvados arvados pipelines cost and compute efficient for production level data. This page will go over Arvados best practices for making your pipeline cost and compute efficient. h2. Pipeline design h3. Choosing the right number of jobs The right number of jobs depends on how versatile you want your pipeline to be. Do you want to output all your intermediate files to keep? How many checkpoints do you want your pipeline to have? One thing to note is that each job must output a collection of files. If you don't want to output files from a command, you should combine multiple commands in a job. You always choose what to upload to keep, so if you don't need files later on, its best to leave it on the compute node. If you want a lot of checkpoints you should have a job for each command. You'll be able to resume/restart work easily if any unexpected interruption happens. Also, you can choose different node types for each command. For example, BWA-mem can scale a lot better than fastqc or varscan, so having a 16 core node for something that doesn't have native multi-threading would be wasteful. h3. Choosing the right number of tasks max_tasks_per_node allows you to choose how many tasks you would like to run on a machine. For example, if you have a lot of small tasks that use 1 core/1GB ram, you can put multiple of those on a bigger machine. For example, 8 tasks on an 8 core machine. If you want to utilize machines better for cost savings, you should use crunchstat-summary to find out the maximum memory/cpu usage for one task, and see if you can fit more than 1 of those on a machine. One warning, however is if you do run out of RAM (some compute nodes can't swap) your process will die with an extraneous error. Sometimes the error is obvious, sometimes its a red herring. h3. How to optimize the number of tasks when you don't have native multithreading Tools like GATK have native multi-threading where if you pass a -t, it will use the correct number of cores on the node. You usually want take advantage of this, and choose the min_cores_per_node that equals your threading parameter. You can use any number of min_tasks_per_node making sure that your tool-threading*min_tasks_per_node is <= min_cores_per_node. Also making sure that your node has enough RAM to allocate to all the tasks. Tools like varscan/freebayes don't have native multi-threading so you need to find a workaround. Generally, these tools have a -L/--intervals to pass in certain loci to work on. If you have a bed file you can split reads on, then you can create a new task per interval. Then, have a job merge the outputs together. h3. Piping between tools or writing to a temporary directory. Creating pipes between tools has shown to sometimes be faster than writing/reading from disk. Feel free to pipe your tools together, for example using subprocess.PIPE in the "python subprocess module":https://docs.python.org/2/library/subprocess.html. Sometimes piping is faster, sometimes it's not. You'll have to try for yourself. h3. Writing to keep There are two ways to write your output collection to keep. Writing straight to keep ( arvados.crunch.TaskOutputDir() ) and staging a file in a temporary directory and then uploading to keep. In general, writing straight to keep will reap more benefits. TaskOutputDir acts like a pipe, so you never have to spend node time on uploading data. One problem though is if your job is dependent on using your output directory as a temporary space for files. If your job uses its output directory for computation, then your job will be trying to compute over a network and could become very slow. That being said, it's very safe for a job to write to a temporary directory then spending compute time uploading to keep. If you have time, it's worth trying both and seeing how much time you save by doing both. Most of the time, writing straight to keep using TaskOutputDir will be the right option, but using a tmpdir is always the safe alternative. h2. Crunchstat Summary Crunchstat-summary is an arvados tool to help choose optimal configurations for arvados jobs and pipeline instances. It helps you choose "runtime_constraints":http://doc.arvados.org/api/schema/Job.html specified in the pipeline template under each job by graphing job statistics. For example: CPU usage, RAM, and Keep network traffic over time. h3. How to install crunchstat-summary <pre> $ git clone https://github.com/curoverse/arvados.git $ cd arvados/tools/crunchstat-summary/ $ python setup.py build $ python setup.py install --user </pre> h3. How to use crunchstat-summary <pre> $ ./bin/crunchstat-summary --help usage: crunchstat-summary [-h] [--job UUID | --pipeline-instance UUID | --log-file LOG_FILE] [--skip-child-jobs] [--format {html,text}] [--verbose] Summarize resource usage of an Arvados Crunch job optional arguments: -h, --help show this help message and exit --job UUID Look up the specified job and read its log data from Keep (or from the Arvados event log, if the job is still running) --pipeline-instance UUID Summarize each component of the given pipeline instance --log-file LOG_FILE Read log data from a regular file --skip-child-jobs Do not include stats from child jobs --format {html,text} Report format --verbose, -v Log more information (once for progress, twice for debug) </pre> There are two ways of using crunchstat-summary: a text view for an overall view of a job or an html page, which graphs usage over time. Case 1: A job that does bwa-aln mapping and converts to bam using samtools. <pre> category metric task_max task_max_rate job_total blkio:202:0 read 310334464 - 913853440 blkio:202:0 write 2567127040 - 7693406208 blkio:202:16 read 8036201472 155118884.01 4538585088 blkio:202:16 write 55502038016 0 0 blkio:202:32 read 2756608 100760.59 6717440 blkio:202:32 write 53570560 0 99514368 cpu cpus 8 - - cpu sys 1592.34 1.17 805.32 cpu user 11061.28 7.98 4620.17 cpu user+sys 12653.62 8.00 5425.49 mem cache 7454289920 - - mem pgmajfault 1859 - 830 mem rss 7965265920 - - mem swap 5537792 - - net:docker0 rx 2023609029 - 2093089079 net:docker0 tx 21404100070 - 49909181906 net:docker0 tx+rx 23427709099 - 52002270985 net:eth0 rx 44750669842 67466325.07 14233805360 net:eth0 tx 2126085781 20171074.09 3670464917 net:eth0 tx+rx 46876755623 67673532.73 17904270277 time elapsed 949 - 1899 # Number of tasks: 3 # Max CPU time spent by a single task: 12653.62s # Max CPU usage in a single interval: 799.88% # Overall CPU usage: 285.70% # Max memory used by a single task: 7.97GB # Max network traffic in a single task: 46.88GB # Max network speed in a single interval: 67.67MB/s # Keep cache miss rate 0.00% # Keep cache utilization 0.00% #!! qr1hi-8i9sb-bzn6hzttfu9cetv max CPU usage was 800% -- try runtime_constraints "min_cores_per_node":8 #!! qr1hi-8i9sb-bzn6hzttfu9cetv max RSS was 7597 MiB -- try runtime_constraints "min_ram_mb_per_node":7782 </pre> !86538baca4ecef099d9fad76ad9c7180.png! Here, you can see the distinct computation steps between the bwa-aln and the samtools step. Since there is a noticeable plateau on CPU usage for both computations, it would be worth trying to run the job on a bigger node. For example, a 16 core node to see if the computation can scale higher than 8 cores. Another thing to note is you can also see the runtime_constraints recommendations. These recommendations are for you to set to ensure the job will be able to call the right node type and run reliably when reproduced. Case study 2: FastQC <pre> category metric task_max task_max_rate job_total blkio:0:0 read 174349211138 65352499.20 174349211138 blkio:0:0 write 0 0 0 cpu cpus 8 - - cpu sys 364.95 0.17 364.95 cpu user 17589.59 6.59 17589.59 cpu user+sys 17954.54 6.72 17954.54 fuseops read 1330241 498.40 1330241 fuseops write 0 0 0 keepcache hit 2655806 1038.00 2655806 keepcache miss 2633 1.60 2633 keepcalls get 2658439 1039.00 2658439 keepcalls put 0 0 0 mem cache 19836608512 - - mem pgmajfault 19 - 19 mem rss 1481367552 - - net:eth0 rx 178321 17798.40 178321 net:eth0 tx 7156 685.00 7156 net:eth0 tx+rx 185477 18483.40 185477 net:keep0 rx 175959092914 107337311.20 175959092914 net:keep0 tx 0 0 0 net:keep0 tx+rx 175959092914 107337311.20 175959092914 time elapsed 3301 - 3301 # Number of tasks: 1 # Max CPU time spent by a single task: 17954.54s # Max CPU usage in a single interval: 672.01% # Overall CPU usage: 543.91% # Max memory used by a single task: 1.48GB # Max network traffic in a single task: 175.96GB # Max network speed in a single interval: 107.36MB/s # Keep cache miss rate 0.10% # Keep cache utilization 99.09% #!! qr1hi-8i9sb-nxqqxravvapt10h max CPU usage was 673% -- try runtime_constraints "min_cores_per_node":7 #!! qr1hi-8i9sb-nxqqxravvapt10h max RSS was 1413 MiB -- try runtime_constraints "min_ram_mb_per_node":1945 </pre> !62222dc72a51c18c15836796e91f3bc7.png! One thing to point out here is "keep_cache utilization":http://doc.arvados.org/api/schema/Job.html, which can be changed using 'keep_cache_mb_per_task'. You can see keep cache utilization at 99.09%, which means its at a good point. You can try increasing this since it is almost at 100%, but it may not yield significant gains. Another thing to note is to look at the CPU usage and keep transfer rate graphs. You should look to see if they ever mirror each other, which is a sign of a cpu bound job, or an i/o bound job. For example, if keep transfer is low but CPU usage is high, then your job is highly dependent on CPU, which means you should upgrade to a higher core node. If CPU usage is low and keep transfer is high, then you may want to increase the keep_cache_mb_per_task in order to be able to compute on more data. h2. Job Optimization h3. Writing to keep There are two ways to write your output collection to keep. Writing straight to keep ( arvados.crunch.TaskOutputDir() ) and staging a file in a temporary directory and then uploading to keep. In general, writing straight to keep will reap more benefits. TaskOutputDir acts like a pipe, so you never have to spend node time on uploading data. One problem though is if your job is dependent on using your output directory as a temporary space for files. If your job uses its output directory for computation, then your job will be trying to compute over a network and could become very slow. That being said, it's very safe for a job to write to a temporary directory then spending compute time uploading to keep. If you have time, it's worth trying both and seeing how much time you save by doing both. Most of the time, writing straight to keep using TaskOutputDir will be the right option, but using a tmpdir is always the safe alternative. h3. Choosing the right number of jobs The right number of jobs depends on how versatile you want your pipeline to be. Do you want to output all your intermediate files to keep? How many checkpoints do you want your pipeline to have? One thing to note is that each job must output a collection of files. If you don't want to output files from a command, you should combine multiple commands in a job. You always choose what to upload to keep, so if you don't need files later on, its best to leave it on the compute node. If you want a lot of checkpoints you should have a job for each command. You'll be able to resume/restart work easily if any unexpected interruption happens. Also, you can choose different node types for each command. For example, BWA-mem can scale a lot better than fastqc or varscan, so having a 16 core node for something that doesn't have native multi-threading would be wasteful. h3. Choosing the right number of tasks max_tasks_per_node allows you to choose how many tasks you would like to run on a machine. For example, if you have a lot of small tasks that use 1 core/1GB ram, you can put multiple of those on a bigger machine. For example, 8 tasks on an 8 core machine. If you want to utilize machines better for cost savings, you should use crunchstat-summary to find out the maximum memory/cpu usage for one task, and see if you can fit more than 1 of those on a machine. One warning, however is if you do run out of RAM (some compute nodes can't swap) your process will die with an extraneous error. Sometimes the error is obvious, sometimes its a red herring. h3. How to optimize the number of tasks when you don't have native multithreading Tools like GATK have native multi-threading where if you pass a -t, it will use the correct number of cores on the node. You usually want take advantage of this, and choose the min_cores_per_node that equals your threading parameter. You can use any number of min_tasks_per_node making sure that your tool-threading*min_tasks_per_node is <= min_cores_per_node. Also making sure that your node has enough RAM to allocate to all the tasks. Tools like varscan/freebayes don't have native multi-threading so you need to find a workaround. Generally, these tools have a -L/--intervals to pass in certain loci to work on. If you have a bed file you can split reads on, then you can create a new task per interval. Then, have a job merge the outputs together. h3. Piping between tools or writing to a temporary directory. Creating pipes between tools has shown to sometimes be faster than writing/reading from disk. Feel free to pipe your tools together, for example using subprocess.PIPE in the "python subprocess module":https://docs.python.org/2/library/subprocess.html. Sometimes piping is faster, sometimes it's not. You'll have to try for yourself.