Project

General

Profile

Pipeline Optimization » History » Version 19

Bryan Cosca, 04/21/2016 03:38 PM

1 1 Bryan Cosca
h1. Pipeline Optimization
2
3 17 Bryan Cosca
h2. Overview
4 1 Bryan Cosca
5 19 Bryan Cosca
This wiki page is designed to help users make their Arvados pipelines cost and compute efficient for production level data. It will teach users how to make the most out of their pipeline to save money or compute time on their clusters.
6 17 Bryan Cosca
7
h2. Pipeline design
8
9
h3. Choosing the right number of jobs
10 1 Bryan Cosca
11 19 Bryan Cosca
Currently, your pipeline may have one script that runs everything from alignment to variant calling. If something unexpected happens and your pipeline fails, you don't want to start from the beginning again, you want to resume from where you left off!
12 1 Bryan Cosca
13 19 Bryan Cosca
This is where the notion of choosing which jobs to create comes into play. The right number of jobs depends on how versatile you want your pipeline to be. Specifically, how many steps do you want your pipeline to have?
14
15
Questions you may ask yourself are: 
16
17 18 Bryan Cosca
Do I want to output all your intermediate files to keep? 
18 1 Bryan Cosca
How many checkpoints do I want my pipeline to have?
19 19 Bryan Cosca
Do I want to do alignment and variant calling in one step? Should I separate them? (usually, no!)
20 1 Bryan Cosca
21 19 Bryan Cosca
Each job must output a collection of files. This is how you can checkpoint your pipeline. The outputs of jobs are always available, so you can reuse them when the next job fails.
22 18 Bryan Cosca
23 19 Bryan Cosca
If you don't want to have a lot of jobs or a lot of outputs, you should combine multiple commands in a job. You can do all your bam realignment/recalibration in one step, and you don't need to output all the intermediate bam files to keep. Keep in mind, 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. One way to do so is to use arvados.current_task().tmpdir to store all your intermediate files, and then only upload what you want to keep.
24
25
If you choose to do multiple computations in a job, you should try piping them together. 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. 
26
27 1 Bryan Cosca
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.
28
29 19 Bryan Cosca
h3. Writing to keep
30 1 Bryan Cosca
31 19 Bryan Cosca
Once you've chosen how to create jobs, you'll need a way to write your outputs to keep.
32 1 Bryan Cosca
33 19 Bryan Cosca
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.
34
35
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.
36
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.
37
38 1 Bryan Cosca
h3. Choosing the right number of tasks
39
40 19 Bryan Cosca
Once you have your job architecture set up, you need to start looking at the core level of the compute node. Specifically, how much compute are you using and how can you take advantage of a machine.
41 1 Bryan Cosca
42 19 Bryan Cosca
If you want multiple computations (tasks) on a node, setting max_tasks_per_node in "runtime_constraints":http://doc.arvados.org/api/schema/Job.html 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.
43 17 Bryan Cosca
44 19 Bryan Cosca
If you know how much RAM you want your job to have, you can set min_ram_mb_per_node in runtime_constraints to ensure that you get a compute node big enough for your computation. Most tools scale up with more RAM, so feel free to test your job on different nodetypes to see what it works on the best.
45 17 Bryan Cosca
46 19 Bryan Cosca
If you have a job that will multithread, you should set num_cores_per_node in runtime_constraints to the number of threads you want. Most clouds have machines up to 16 threads, so it would be worth testing on multiple nodetypes to see what works best for you.
47 17 Bryan Cosca
48 19 Bryan Cosca
If you have a job that's very i/o intensive, its worth it to change keep_cache_mb_per_task in runtime_constraints 
49 17 Bryan Cosca
50
51 19 Bryan Cosca
h3. How to optimize the number of tasks when you don't have native multithreading
52 17 Bryan Cosca
53 19 Bryan Cosca
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.
54
55
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.
56
57 2 Bryan Cosca
h2. Crunchstat Summary
58 16 Bryan Cosca
59 3 Bryan Cosca
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.
60 1 Bryan Cosca
61 2 Bryan Cosca
h3. How to install crunchstat-summary
62 3 Bryan Cosca
63
<pre>
64
$ git clone https://github.com/curoverse/arvados.git
65
$ cd arvados/tools/crunchstat-summary/
66
$ python setup.py build
67
$ python setup.py install --user
68
</pre>
69 1 Bryan Cosca
70
h3. How to use crunchstat-summary
71 3 Bryan Cosca
72
<pre>
73
$ ./bin/crunchstat-summary --help
74
usage: crunchstat-summary [-h]
75
                          [--job UUID | --pipeline-instance UUID | --log-file LOG_FILE]
76
                          [--skip-child-jobs] [--format {html,text}]
77
                          [--verbose]
78
79
Summarize resource usage of an Arvados Crunch job
80
81
optional arguments:
82
  -h, --help            show this help message and exit
83
  --job UUID            Look up the specified job and read its log data from
84
                        Keep (or from the Arvados event log, if the job is
85
                        still running)
86 1 Bryan Cosca
  --pipeline-instance UUID
87 3 Bryan Cosca
                        Summarize each component of the given pipeline
88
                        instance
89
  --log-file LOG_FILE   Read log data from a regular file
90
  --skip-child-jobs     Do not include stats from child jobs
91
  --format {html,text}  Report format
92
  --verbose, -v         Log more information (once for progress, twice for
93
                        debug)
94
</pre>
95 1 Bryan Cosca
96 16 Bryan Cosca
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.
97 14 Bryan Cosca
98 16 Bryan Cosca
Case 1: A job that does bwa-aln mapping and converts to bam using samtools.
99 8 Bryan Cosca
100
<pre>
101 19 Bryan Cosca
$ ~/arvados/tools/crunchstat-summary/bin/crunchstat-summary --format text --job qr1hi-8i9sb-bzn6hzttfu9cetv
102 8 Bryan Cosca
category        metric  task_max        task_max_rate   job_total
103
blkio:202:0     read    310334464       -       913853440
104
blkio:202:0     write   2567127040      -       7693406208
105
blkio:202:16    read    8036201472      155118884.01    4538585088
106
blkio:202:16    write   55502038016     0       0
107
blkio:202:32    read    2756608 100760.59       6717440
108
blkio:202:32    write   53570560        0       99514368
109
cpu     cpus    8       -       -
110
cpu     sys     1592.34 1.17    805.32
111
cpu     user    11061.28        7.98    4620.17
112
cpu     user+sys        12653.62        8.00    5425.49
113
mem     cache   7454289920      -       -
114
mem     pgmajfault      1859    -       830
115
mem     rss     7965265920      -       -
116
mem     swap    5537792 -       -
117
net:docker0     rx      2023609029      -       2093089079
118
net:docker0     tx      21404100070     -       49909181906
119
net:docker0     tx+rx   23427709099     -       52002270985
120
net:eth0        rx      44750669842     67466325.07     14233805360
121
net:eth0        tx      2126085781      20171074.09     3670464917
122 1 Bryan Cosca
net:eth0        tx+rx   46876755623     67673532.73     17904270277
123
time    elapsed 949     -       1899
124 8 Bryan Cosca
# Number of tasks: 3
125
# Max CPU time spent by a single task: 12653.62s
126
# Max CPU usage in a single interval: 799.88%
127
# Overall CPU usage: 285.70%
128
# Max memory used by a single task: 7.97GB
129
# Max network traffic in a single task: 46.88GB
130
# Max network speed in a single interval: 67.67MB/s
131
# Keep cache miss rate 0.00%
132
# Keep cache utilization 0.00%
133 1 Bryan Cosca
#!! qr1hi-8i9sb-bzn6hzttfu9cetv max CPU usage was 800% -- try runtime_constraints "min_cores_per_node":8
134 8 Bryan Cosca
#!! qr1hi-8i9sb-bzn6hzttfu9cetv max RSS was 7597 MiB -- try runtime_constraints "min_ram_mb_per_node":7782
135
</pre>
136
137 19 Bryan Cosca
$ ~/arvados/tools/crunchstat-summary/bin/crunchstat-summary --format html --job qr1hi-8i9sb-bzn6hzttfu9cetv > qr1hi-8i9sb-bzn6hzttfu9cetv.html
138
139 1 Bryan Cosca
!86538baca4ecef099d9fad76ad9c7180.png!
140
141 16 Bryan Cosca
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. 
142 1 Bryan Cosca
143 16 Bryan Cosca
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.
144
145 12 Bryan Cosca
Case study 2: FastQC
146
147
<pre>
148 19 Bryan Cosca
$ ~/arvados/tools/crunchstat-summary/bin/crunchstat-summary --format text --job qr1hi-8i9sb-nxqqxravvapt10h
149 12 Bryan Cosca
category	metric	task_max	task_max_rate	job_total
150
blkio:0:0	read	174349211138	65352499.20	174349211138
151
blkio:0:0	write	0	0	0
152
cpu	cpus	8	-	-
153
cpu	sys	364.95	0.17	364.95
154
cpu	user	17589.59	6.59	17589.59
155 1 Bryan Cosca
cpu	user+sys	17954.54	6.72	17954.54
156
fuseops	read	1330241	498.40	1330241
157 16 Bryan Cosca
fuseops	write	0	0	0
158 1 Bryan Cosca
keepcache	hit	2655806	1038.00	2655806
159
keepcache	miss	2633	1.60	2633
160
keepcalls	get	2658439	1039.00	2658439
161 16 Bryan Cosca
keepcalls	put	0	0	0
162 1 Bryan Cosca
mem	cache	19836608512	-	-
163 16 Bryan Cosca
mem	pgmajfault	19	-	19
164 1 Bryan Cosca
mem	rss	1481367552	-	-
165 16 Bryan Cosca
net:eth0	rx	178321	17798.40	178321
166
net:eth0	tx	7156	685.00	7156
167 11 Bryan Cosca
net:eth0	tx+rx	185477	18483.40	185477
168 16 Bryan Cosca
net:keep0	rx	175959092914	107337311.20	175959092914
169 1 Bryan Cosca
net:keep0	tx	0	0	0
170
net:keep0	tx+rx	175959092914	107337311.20	175959092914
171
time	elapsed	3301	-	3301
172 16 Bryan Cosca
# Number of tasks: 1
173 1 Bryan Cosca
# Max CPU time spent by a single task: 17954.54s
174 16 Bryan Cosca
# Max CPU usage in a single interval: 672.01%
175 1 Bryan Cosca
# Overall CPU usage: 543.91%
176 16 Bryan Cosca
# Max memory used by a single task: 1.48GB
177 1 Bryan Cosca
# Max network traffic in a single task: 175.96GB
178 16 Bryan Cosca
# Max network speed in a single interval: 107.36MB/s
179
# Keep cache miss rate 0.10%
180 11 Bryan Cosca
# Keep cache utilization 99.09%
181
#!! qr1hi-8i9sb-nxqqxravvapt10h max CPU usage was 673% -- try runtime_constraints "min_cores_per_node":7
182 5 Bryan Cosca
#!! qr1hi-8i9sb-nxqqxravvapt10h max RSS was 1413 MiB -- try runtime_constraints "min_ram_mb_per_node":1945
183
</pre>
184 19 Bryan Cosca
185
~/arvados/tools/crunchstat-summary/bin/crunchstat-summary --format html --job qr1hi-8i9sb-nxqqxravvapt10h > qr1hi-8i9sb-nxqqxravvapt10h.html
186 16 Bryan Cosca
187 1 Bryan Cosca
!62222dc72a51c18c15836796e91f3bc7.png!
188 16 Bryan Cosca
189 5 Bryan Cosca
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.
190 11 Bryan Cosca
191 1 Bryan Cosca
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.