% Savio introductory training: Basic usage of the Berkeley Savio high-performance computing cluster % February 11, 2021 % Clint Hamilton, James Duncan and Chris Paciorek
-
Lots of other events at Berkeley this week during Love Data Week.
-
Cloud Computing Meetup (monthly)
-
Women in Data Science Berkeley (March 8-12, 2021)
-
Looking for researchers working with sensitive data as we are building tools and services to support that work. Get in touch for more information.
-
Securing Research Data Working Group (monthly)
-
Research IT is hiring graduate students as domain consultants. See flyers or talk to one of us.
We'll do this mostly as a demonstration. We encourage you to login to your account and try out the various examples yourself as we go through them.
Much of this material is based on the extensive Savio documention we have prepared and continue to prepare, available at https://docs-research-it.berkeley.edu/services/high-performance-computing/.
The materials for this tutorial are available using git at the short URL (tinyurl.com/brc-feb21), the GitHub URL (https://github.com/ucb-rit/savio-training-intro-spring-2021), or simply as a zip file.
This training session will cover the following topics:
- Introductory content
- Basic parallel computing concepts
- High level overview of system
- System capabilities and hardware
- Getting access to the system - FCA, condo, ICA
- Login nodes, compute nodes, and DTN nodes
- Savio computing nodes
- Disk space options (home, scratch, group, condo storage)
- Logging in, data transfer, and software
- Logging in
- Data transfer
- SCP/SFTP
- Globus
- Box & bDrive (Google drive)
- Software modules
- Submitting and monitoring jobs
- Acounts and partitions
- Basic job submission
- Parallel jobs
- Interactive jobs
- Low-priority queue
- HTC jobs
- Monitoring jobs and cluster status
- Basic use of standard software: Python and R
- Jupyter notebooks using OOD
- Parallelization in Python with ipyparallel
- Parallelization in R with foreach
- Dask for parallelization in Python
- More information
- How to get additional help
- Upcoming events
- Who is Savio for?
- Types of computation (Flynn's taxonomy)
- Single command, Single Data
- Single command, Multiple data
- Multiple command, Multiple data
- (Multiple Instruction, Single data)
- Savio is (generally) best for cases 2 & 3
- What is Savio?
- In layman's terms:
- A collection of really powerful computers (nodes)
- Some really big, fast hard drives
- In layman's terms:
- Two types of parallel Computing
- Shared memory (OpenMP)
- All computation on the same node
- Can have shared objects in ram
- Distributed memory (MPI)
- Computation on multiple nodes
- Special attention to passing information between nodes
- Shared memory (OpenMP)
- All regular Berkeley faculty can request 300,000 service units (roughly core-hours) per year through the Faculty Computing Allowance (FCA)
- Researchers can also purchase nodes for their own priority access and gain access to the shared Savio infrastructure and to the ability to burst to additional nodes through the condo cluster program
- Instructors can request an Instructional Computing Allowance (ICA).
Faculty/principal investigators can allow researchers working with them to get user accounts with access to the FCA or condo resources available to the faculty member.
- Savio is a 600-node, >15,000-core Linux cluster rated at nearly 540 peak teraFLOPS.
- about 40% of the compute nodes provided by the institution for general access
- about 60% compute nodes contributed by researchers in the Condo program
Savio is a Linux cluster - by cluster we mean a set of computers networked together
- Savio has 3 kinds of nodes:
- Login nodes
- Data transfer nodes
- Compute nodes
- Login nodes
- Used to access the system when loggin in
- For login and non-intensive interactive work such as:
- job submission and monitoring
- basic compilation
- managing your disk space
- Data transfer nodes
- For transferring data to/from Savio
- This is a notable difference from many other cluster
- Login node:
hpc.brc.berkeley.edu
- Data transfer node:
dtn.brc.berkeley.edu
- Some applications may look for SFTP via login node
- Login node:
- Note: you can access your files on the system from any of the computers
- Compute nodes
- For computational tasks
- Your work might use parallelization to do computation on more than one CPU
- You can also run "serial" jobs that use a single CPU
-
There are multiple types of computation nodes with different hardware specifications (see the Hardware Configuration page).
-
The nodes are divided into several pools, called partitions
-
These partitions have different restrictions and costs associated with them
-
Any job you submit must be submitted to a partition to which you have access.
- You have access to the multiple kinds of disk space, described here in the Storing Data page.
- There are 3 directories:
/global/home/users/SAVIO_USERNAME
- 10 GB per user, backed up
/global/home/groups/SAVIO_GROUPNAME
- Per group: 30 GB for FCA, 200 GB for Condo
/global/scratch/SAVIO_USERNAME
- 12 TB per user
- Connected via Infiniband (very fast)
- Primary data storage during computation
- All 3 are available from any of the nodes and changes to files on one node will be seen on all the other nodes
- Large amounts of disk space is available for purchase from the condo storage offering.
- The minimum purchase is about $6,200, which provides roughly 42 TB for five years
- When reading/writing data to/from disk put the data in your scratch space at
/global/scratch/SAVIO_USERNAME
- The system is set up so that disk access for all users is optimized when users are doing input/output (I/O) off of scratch rather than off of their home directories
- Doing I/O with files on your home directory can impact the ability of others to access their files on the filesystem
- Savio (and AEoD) is certified for moderately sensitive data
- P2, P3 (formerly PL1) and NIH dbGap (non-"notice-triggering" data).
- PIs/faculty must request a P2/P3 project alongside requests for a new FCA/condo allocation
- Existing projects can't be converted to P2/P3 projects.
- BRC has a new platform for highly sensitive data (P4) called SRDC.
More info is available in the slides from Dec. 2019 "Working with Sensitive + Protected Data" workshop: https://tinyurl.com/srdc-dec2019
- To login, you need to have software on your own machine that gives you access to the SSH command
- These come built-in with Mac (see
Applications -> Utilities -> Terminal
). - For Windows, you can use Powershell (or Command Prompt)
- Other application such as MobaXterm may offer more functionality
- These come built-in with Mac (see
- You also need to set up your smartphone or tablet with Google Authenticator to generate one-time passwords for you.
- Here are instructions for doing this setup, and for logging in.
Then to login:
-
Then enter XXXXXYYYYYY where XXXXXX is your PIN and YYYYYY is the one-time password. YYYYYY will be shown when you open your Google authenticator app on your phone/tablet.
-
One can then navigate around and get information using standard UNIX commands such as
ls
,cd
,du
,df
, etc.- There is a lot of material online about using the UNIX command line
- Also called the shell; 'bash' is one common variation
- Here is a basic tutorial.
- There is a lot of material online about using the UNIX command line
If you want to be able to open programs with graphical user interfaces:
ssh -Y [email protected]
- To display the graphical windows on your local machine, you'll need X server software on your own machine to manage the graphical windows
- For Windows, your options include MobaXterm, eXceed, or Xming
- For Mac, there is XQuartz
You are welcome to edit your files on Savio (rather than copying files back and forth from your laptop and editing them on your laptop). To do so you'll need to use some sort of editor. Savio has vim
, emacs
, and nano
installed. Just start the editor from a login node.
## To use vim
vim myfile.txt
## To use emacs
emacs myfile.txt
## To use nano
module load nano
nano myfile.txt
To do any work on the system, you'll usually need to transfer files (data files, code files, etc.) to the Savio filesystem, either into your home directory, your scratch directory or a group directory.
And once you're done with a computation, you'll generally need to transfer files back to another place (e.g., your laptop).
Let's see how we would transfer files/data to/from Savio using a few different approaches.
-
The most common command line protocol for file transfer is SCP
-
You need to use the Savio data transfer node,
dtn.brc.berkeley.edu
. -
The example file
bayArea.csv
is too large to store on Github; you can obtain it here. -
SCP is supported in terminal for Mac/Linux and in Powershell/Command Prompt in Windows
# to Savio, while on your local machine
scp bayArea.csv [email protected]:~/.
scp bayArea.csv [email protected]:~/data/newName.csv
scp bayArea.csv [email protected]:/global/scratch/paciorek/.
# from Savio, while on your local machine
scp [email protected]:~/data/newName.csv ~/Documents/.
If you can ssh to your local machine or want to transfer files to other systems on to which you can ssh, you can login to the dtn node to execute the scp commands:
ssh [email protected]
[SAVIO_USERNAME@dtn ~]$ scp ~/file.csv [email protected]:~/data/.
If you're already connected to a Savio login node, you can use ssh dtn
to login to the dtn.
Pro tip: You can package multiple files (including directory structure) together using tar
tar -cvzf files.tgz dir_to_zip
# to untar later:
tar -xvzf files.tgz
- Another common method for file transfer is SFTP
- A multi-platform program for doing transfers via SFTP is FileZilla.
- After logging in to most SFTP applications, you'll see windows for the Savio filesystem and your local filesystem on your machine. You can drag files back and forth.
-
You can use Globus Connect to transfer data data to/from Savio (and between other resources) quickly and unattended
- This is a better choice for large transfers
- Here are some instructions.
-
Globus transfers data between endpoints
- Possible endpoints include
- Savio
- your laptop or desktop
- Other clusters like NERSC and XSEDE
- Possible endpoints include
-
If you are transferring to/from your laptop, you'll need
- Globus Connect Personal set up,
- your machine established as an endpoint, and
- Globus Connect Pesonal actively running on your machine. At that point you can proceed as below.
-
Savio's endpoint is named
ucb#brc
.
-
To transfer files, you open Globus at globus.org and authenticate to the endpoints you want to transfer between.
- This means that you only need to authenticate once, whereas you might need to authenticate multiple times with scp and sftp.
- You can then start a transfer and it will proceed in the background, including restarting if interrupted.
-
Globus also provides a command line interface that will allow you to do transfers programmatically
- Thus a transfer could be embedded in a workflow script.
-
Box and bDrive (the Cal branded Google Drive) both provide free, secured, and encrypted content storage of files to Berkeley affiliates
- bDrive provides unlimited storage (for now), Box is currently also available
- bDrive has a maximum file size of 5Tb, Box has a maximum file size of 15 Gb
- They are both good options for backup and long-term storage
- However, since the UC Berkeley contract with Box ends in 2023 we suggest new users opt for bDrive
- bDrive provides unlimited storage (for now), Box is currently also available
-
You can interact with both services via web browser, and both services provide a desktop app you can use to move and sync files between your computer and the cloud.
-
BRC is working on making Globus available for transfer to/from Box and bDrive
- We hope this will be available in the near future
For more ambitious users, Box has a Python-based SDK that can be used to write scripts for file transfers. For more information on how to do this, check out the BoxAuthenticationBootstrap.ipynb
and TransferFilesFromBoxToSavioScratch.ipynb
from BRC's cyberinfrastructure engineer on GitHub
rclone is a command line program that you can use to sync files between both services and Savio. You can read instructions for using rclone on Savio with Box or bDrive here.
Briefly the steps to set up rclone on Savio to interact with Box are as follows:
- Configuration (on dtn):
rclone config
- Use auto config? -> n
- For Box: install rclone on your PC, then run
rclone authorize "box"
- Paste the link into your browser and log in to your CalNet account
- Copy the authentication token and paste into the
rclone config
prompt on Savio
Finally you can set up special purpose accounts so files are owned at a project level rather than by individuals.
rclone basics:
- Switch to DTN before using if on login node
- Use command
ssh dtn
- If using rclone on another node You need to load rclone before use
- Run command
module load rclone
- Run command
- Use command
- All rclone commands begin with
rclone
and are then followed by a commands- The commands are different from bash (i.e.,
cp
in bash vs.copy
in rclone)
- The commands are different from bash (i.e.,
- To reference a file on the remote you add configured remote name followed by a colon followed by the file path
- For example
clint_bdrive:project_folder
- To access the main folder leave nothing after the colon (e.g.,
clint_bdrive:
)
- For example
- For more tips and tricks see our docs
rclone example:
rclone listremotes # Lists configured remotes.
rclone lsd remote_name: # Lists directories, but not files. Note the trailing colon.
rclone size remote_name:home # Prints size and number of objects in remote "home" directory. This can take a very long time when tallying Tbs of files.
rclone copy /global/home/users/hannsode remote_name:savio_home/hannsode # Copies my entire home directory to a new directory on the remote.
rclone copy /global/scratch/hannsode/genomes remote_name:genome_sequences # Copies entire directory contents to a dirctory on the remote with a new name.
A lot of software is available on Savio but needs to be loaded from the relevant software module before you can use it. (We do this not to confuse you but to avoid clashes between incompatible software and allow multiple versions of a piece of software to co-exist on the system.)
module list # what's loaded?
module avail # what's available
One thing that tricks people is that the modules are arranged in a hierarchical (nested) fashion, so you only see some of the modules as being available after you load the parent module (e.g., MKL, FFT, and HDF5/NetCDF software is nested within the gcc module). Here's how we see and load MPI.
module load openmpi # this fails if gcc not yet loaded
module load gcc
module avail
module load openmpi
Note that a variety of Python packages are available simply by loading the python module. For R this is not the case, but you can load the r-packages module.
All computations are done by submitting jobs to the scheduling software that manages jobs on the cluster, called SLURM.
Why is this necessary? Otherwise your jobs would be slowed down by other people's jobs running on the same node. This also allows everyone to fairly share Savio.
The basic workflow is:
- login to Savio; you'll end up on one of the login nodes in your home directory
- use
cd
to go to the directory from which you want to submit the job - submit the job using
sbatch
(or an interactive job usingsrun
, discussed later)- when your job starts, the working directory will be the one from which the job was submitted
- the job will be running on a compute node, not the login node
When submitting a job, the main things you need to indicate are the project account you are using and the partition. Note that there is a default value for the project account, but if you have access to multiple accounts such as an FCA and a condo, it's good practice to specify it.
You can see what accounts you have access to and which partitions within those accounts as follows:
sacctmgr -p show associations user=$USER
Here's an example of the output for a user who has access to an FCA, a condo, and a special partner account:
Cluster|Account|User|Partition|Share|GrpJobs|GrpTRES|GrpSubmit|GrpWall|GrpTRESMins|MaxJobs|MaxTRES|MaxTRESPerNode|MaxSubmit|MaxWall|MaxTRESMins|QOS|Def QOS|GrpTRESRunMins|
brc|co_stat|paciorek|savio2_1080ti|1||||||||||||savio_lowprio|savio_lowprio||
brc|co_stat|paciorek|savio2_knl|1||||||||||||savio_lowprio|savio_lowprio||
brc|co_stat|paciorek|savio2_bigmem|1||||||||||||savio_lowprio|savio_lowprio||
brc|co_stat|paciorek|savio2_gpu|1||||||||||||savio_lowprio,stat_gpu2_normal|stat_gpu2_normal||
brc|co_stat|paciorek|savio2_htc|1||||||||||||savio_lowprio|savio_lowprio||
brc|co_stat|paciorek|savio|1||||||||||||savio_lowprio|savio_lowprio||
brc|co_stat|paciorek|savio_bigmem|1||||||||||||savio_lowprio|savio_lowprio||
brc|co_stat|paciorek|savio2|1||||||||||||savio_lowprio,stat_savio2_normal|stat_savio2_normal||
brc|fc_paciorek|paciorek|savio2_1080ti|1||||||||||||savio_debug,savio_normal|savio_normal||
brc|fc_paciorek|paciorek|savio2_knl|1||||||||||||savio_debug,savio_normal|savio_normal||
brc|fc_paciorek|paciorek|savio2_gpu|1||||||||||||savio_debug,savio_normal|savio_normal||
brc|fc_paciorek|paciorek|savio2_htc|1||||||||||||savio_debug,savio_long,savio_normal|savio_normal||
brc|fc_paciorek|paciorek|savio2_bigmem|1||||||||||||savio_debug,savio_normal|savio_normal||
brc|fc_paciorek|paciorek|savio2|1||||||||||||savio_debug,savio_normal|savio_normal||
brc|fc_paciorek|paciorek|savio|1||||||||||||savio_debug,savio_normal|savio_normal||
brc|fc_paciorek|paciorek|savio_bigmem|1||||||||||||savio_debug,savio_normal|savio_normal||
If you are part of a condo, you'll notice that you have low-priority access to certain partitions. For example, Chris is part of the statistics condo co_stat, which owns some savio2 nodes and savio2_gpu nodes and therefore has normal access to those, but he can also burst beyond the condo and use other partitions at low-priority (see below).
In contrast, through his FCA, he has access to the savio, savio2, big memory, HTC, and GPU partitions all at normal priority.
Let's see how to submit a simple job. If your job will only use the resources on a single node, you can do the following.
Here's an example job script that I'll run. You'll need to modify the --account value and possibly the --partition value.
#!/bin/bash
# Job name:
#SBATCH --job-name=test
#
# Account:
#SBATCH --account=fc_brownlab
#
# Partition:
#SBATCH --partition=savio2
#
# Wall clock limit (5 minutes here):
#SBATCH --time=00:05:00
#
## Command(s) to run:
module load python/3.6
python calc.py >& calc.out
Now let's submit and monitor the job:
sbatch job.sh
squeue -j <JOB_ID>
wwall -j <JOB_ID>
After a job has completed (or been terminated/cancelled), you can review the maximum memory used via the sacct command.
sacct -j <JOB_ID> --format=JobID,JobName,MaxRSS,Elapsed
MaxRSS will show the maximum amount of memory that the job used in kilobytes.
You can also login to the node where you are running and use commands like top and ps:
srun --jobid=<JOB_ID> --pty /bin/bash
Note that except for the savio2_htc and savio2_gpu partitions, all jobs are given exclusive access to the entire node or nodes assigned to the job (and your account is charged for all of the cores on the node(s)).
If you are submitting a job that uses multiple nodes, you'll need to carefully specify the resources you need. The key flags for use in your job script are:
--nodes
(or-N
): indicates the number of nodes to use--ntasks-per-node
: indicates the number of tasks (i.e., processes) one wants to run on each node--cpus-per-task
(or-c
): indicates the number of cpus to be used for each task
In addition, in some cases it can make sense to use the --ntasks
(or -n
) option to indicate the total number of tasks and let the scheduler determine how many nodes and tasks per node are needed. In general --cpus-per-task
will be one except when running threaded code.
Here's an example job script for a job that uses MPI for parallelizing over multiple nodes:
#!/bin/bash
# Job name:
#SBATCH --job-name=test
#
# Account:
#SBATCH --account=account_name
#
# Partition:
#SBATCH --partition=partition_name
#
# Number of MPI tasks needed for use case (example):
#SBATCH --ntasks=40
#
# Processors per task:
#SBATCH --cpus-per-task=1
#
# Wall clock limit:
#SBATCH --time=00:00:30
#
## Command(s) to run (example):
module load intel openmpi
mpirun ./a.out
When you write your code, you may need to specify information about the number of cores to use. SLURM will provide a variety of variables that you can use in your code so that it adapts to the resources you have requested rather than being hard-coded.
Here are some of the variables that may be useful: SLURM_NTASKS, SLURM_CPUS_PER_TASK, SLURM_NODELIST, SLURM_NNODES.
Some common paradigms are:
- 1 node, many CPUs
- openMP/threaded jobs - 1 task, c CPUs for the task
- Python/R/GNU parallel - many tasks, 1 per CPU at any given time
- many nodes, many CPUs
- MPI jobs that use 1 CPU per task for each of n tasks, spread across multiple nodes
- Python/R/GNU parallel - many tasks, 1 per CPU at any given time
- hybrid jobs that use c CPUs for each of n tasks
- e.g., MPI+threaded code
There are lots more examples of job submission scripts for different kinds of parallelization (multi-node (MPI), multi-core (openMP), hybrid, etc.) here.
You can also do work interactively. This simply moves you from a login node to a compute node.
srun -A fc_brownlab -p savio2 --nodes=1 -t 10:0 --pty bash
# note that you end up in the same working directory as when you submitted the job
# now execute on the compute node:
env | grep SLURM
module load matlab
matlab -nodesktop -nodisplay
To end your interactive session (and prevent accrual of additional charges to your FCA), simply enter exit
in the terminal session.
NOTE: you are charged for the entire node when running interactive jobs (as with batch jobs) except in the savio2_htc and various GPU partitions.
If you are running a graphical interface, we recommend you use Savio's remote desktop service on our visualization node, as described here.
Condo users have access to the broader compute resource that is limited only by the size of partitions, under the savio_lowprio QoS (queue). However this QoS does not get a priority as high as the general QoSs, such as savio_normal and savio_debug, or all the condo QoSs, and it is subject to preemption when all the other QoSs become busy.
More details can be found in the Low Priority Jobs section of the user guide.
Suppose I wanted to burst beyond the Statistics condo to run on 20 nodes. I'll illustrate here with an interactive job though usually this would be for a batch job.
First I'll see if there are that many nodes even available.
sinfo -p savio2
srun -A co_stat -p savio2 --qos=savio_lowprio --nodes=20 -t 10:00 --pty bash
## now look at environment variables to see my job can access 20 nodes:
env | grep SLURM
There is a partition called the HTC partition that allows you to request cores individually rather than an entire node at a time. The nodes in this partition are faster than the other nodes. Here is an example SLURM script:
#!/bin/bash
# Job name:
#SBATCH --job-name=test
#
# Account:
#SBATCH --account=account_name
#
# Partition:
#SBATCH --partition=savio2_htc
#
# Processors per task:
#SBATCH --cpus-per-task=2
#
# Wall clock limit -- 10 minutes
#SBATCH --time=00:10:00
#
## Command(s) to run (example):
module load python
python calc.py >& calc.out
One can run jobs up to 10 days (using four or fewer cores) in this partition if you include --qos=savio_long
.
You may have many serial jobs to run. It may be more cost-effective to collect those jobs together and run them across multiple cores on one or more nodes.
Here are some options:
- using GNU parallel to run many computational tasks (e.g., thousands of simulations, scanning tens of thousands of parameter values, etc.) as part of single Savio job submission
- using single-node parallelism and multiple-node parallelism in Python, R, and MATLAB
- parallel R tools such as future, foreach, parLapply, and mclapply
- parallel Python tools such as ipyparallel, Dask, and ray
- parallel functionality in MATLAB through parfor
The basic command for seeing what is running on the system is squeue
:
squeue
squeue -u $USER
squeue -A co_stat
To see what nodes are available in a given partition:
sinfo -p savio
sinfo -p savio2_gpu
You can cancel a job with scancel
.
scancel <YOUR_JOB_ID>
For more information on cores, QoS, and additional (e.g., GPU) resources, here's some syntax:
squeue -o "%.7i %.12P %.20j %.8u %.2t %.9M %.5C %.8r %.3D %.20R %.8p %.20q %b"
We provide some tips about monitoring your jobs.
If you'd like to see how much of an FCA has been used:
check_usage.sh -a fc_cuore
The new sq
tool on Savio provides a bit more user-friendly way to understand why your job isn't running yet. Here's the basic usage:
# should be loaded by default, but if it isn't:
module load sq
sq
Showing results for user jpduncan
Currently 0 running jobs and 1 pending job (most recent job first):
+---------|------|-------------|-----------|--------------|------|---------|-----------+
| Job ID | Name | Account | Nodes | QOS | Time | State | Reason |
+---------|------|-------------|-----------|--------------|------|---------|-----------+
| 7510375 | test | fc_brownlab | 1x savio2 | savio_normal | 0:00 | PENDING | Resources |
+---------|------|-------------|-----------|--------------|------|---------|-----------+
7510375:
This job is scheduled to run after 21 higher priority jobs.
Estimated start time: N/A
To get scheduled sooner, you can try reducing wall clock time as appropriate.
Recent jobs (most recent job first):
+---------|------|-------------|-----------|----------|---------------------|-----------+
| Job ID | Name | Account | Nodes | Elapsed | End | State |
+---------|------|-------------|-----------|----------|---------------------|-----------+
| 7509474 | test | fc_brownlab | 1x savio2 | 00:00:16 | 2021-02-09 23:47:45 | COMPLETED |
+---------|------|-------------|-----------|----------|---------------------|-----------+
7509474:
- This job ran for a very short amount of time (0:00:16). You may want to check that the output was correct or if it exited because of a problem.
To see another user's jobs:
sq -u paciorek
The -a
flag shows current and past jobs together, the -q
flag suppresses messages about job issues, and the -n
flag sets the limit on the number of jobs to show in the output (default = 8).
sq -u paciorek -aq -n 10
Showing results for user paciorek
Recent jobs (most recent job first):
+-----------|------|-------------|-----------|------------|---------------------|-----------+
| Job ID | Name | Account | Nodes | Elapsed | End | State |
+-----------|------|-------------|-----------|------------|---------------------|-----------+
| 7487633.1 | ray | co_stat | 1x | 1-20:19:03 | Unknown | RUNNING |
| 7487633.0 | ray | co_stat | 1x | 1-20:19:08 | Unknown | RUNNING |
| 7487633 | test | co_stat | 2x savio2 | 1-20:19:12 | Unknown | RUNNING |
| 7487879 | bash | ac_scsguest | 1x savio | 00:00:27 | 2021-02-08 14:54:19 | COMPLETED |
| 7487633.2 | bash | co_stat | 2x | 00:00:34 | 2021-02-08 14:53:38 | FAILED |
| 7487515 | test | co_stat | 2x savio2 | 00:04:53 | 2021-02-08 14:22:17 | CANCELLED |
| 7487515.1 | ray | co_stat | 1x | 00:00:06 | 2021-02-08 14:17:39 | FAILED |
| 7487515.0 | ray | co_stat | 1x | 00:00:05 | 2021-02-08 14:17:33 | FAILED |
| 7473988 | test | co_stat | 2x savio2 | 3-00:00:16 | 2021-02-08 13:33:40 | TIMEOUT |
| 7473989 | test | ac_scsguest | 2x savio | 2-22:30:11 | 2021-02-08 11:47:54 | CANCELLED |
+-----------|------|-------------|-----------|------------|---------------------|-----------+
For help with sq
:
sq -h
To learn more, see our page on understanding when your jobs will run.
Savio now has an Open OnDemand portal, allowing users to launch Jupyter notebooks and RStudio servers for interactive debugging and batch computing. (Note: this supersedes the JupyterHub instance on Savio, which will soon become inactive).
Let's see a brief demo of a Jupyter notebook:
- Connect to ood.brc.berkeley.edu
- Login as usual with a one-time password
- Click on the "Interactive Apps" drop-down menu and select how to run your notebook (either "compute in batch queues" or "non-batch for exploration, debugging")
- Specify your "SLURM Project/Account Name"
- Start up a notebook
Alternatively, one can run a Jupyter notebook through the visualization node.
You can also run parallel computations via an IPython notebook.
Let's see a basic example of doing an analysis in Python across multiple cores on multiple nodes. We'll use the airline departure data in bayArea.csv.
Here we'll use IPython for parallel computing. The example is a bit contrived in that a lot of the time is spent moving data around rather than doing computation, but it should illustrate how to do a few things.
First we'll user-install a Python package called statsmodels
to get a more recent version:
cp bayArea.csv /global/scratch/jpduncan/. # remember to do I/O off scratch
# install Python package
module unload python
module load python/3.6
# the system version of statsmodels isn't recent enough
pip install --user statsmodels --upgrade
Now we'll start up an interactive session, though often this sort of thing would be done via a batch job.
srun -A fc_brownlab -p savio2 --nodes=2 --ntasks-per-node=24 -t 30:0 --pty bash
Now we'll start up a cluster using IPython's parallel tools. To do this across multiple nodes within a SLURM job, it goes like this:
module load python/3.6 gcc openmpi
ipcontroller --ip='*' &
sleep 30
## The next line starts one worker per SLURM task (which should equal the number of cores)
srun ipengine &
sleep 45 # wait until all engines have successfully started
cd /global/scratch/jpduncan
ipython
If we were doing this on a single node, we could start everything up in a single call to ipcluster:
module load python/3.6
ipcluster start -n $SLURM_CPUS_ON_NODE &
ipython
Here's our Python code (also found in parallel.py) for doing an analysis across multiple strata/subsets of the dataset in parallel. Note that the 'load_balanced_view' business is so that the computations are done in a load-balanced fashion, which is important for tasks that take different amounts of time to complete.
from ipyparallel import Client
c = Client()
c.ids
dview = c[:] # use all engines
dview.block = True # wait for the result before returning
dview.apply(lambda : "Hello, World")
lview = c.load_balanced_view()
lview.block = True
import pandas
dat = pandas.read_csv('bayArea.csv', header = None, encoding = 'latin1')
dat.columns = ('Year','Month','DayofMonth','DayOfWeek','DepTime',
'CRSDepTime','ArrTime','CRSArrTime','UniqueCarrier','FlightNum',
'TailNum','ActualElapsedTime','CRSElapsedTime','AirTime','ArrDelay',
'DepDelay','Origin','Dest','Distance','TaxiIn','TaxiOut','Cancelled',
'CancellationCode','Diverted','CarrierDelay','WeatherDelay',
'NASDelay','SecurityDelay','LateAircraftDelay')
dview.execute('import statsmodels.api as sm')
dat2 = dat.reindex(columns = ['DepDelay','Year','Dest','Origin'])
dests = dat2.Dest.unique()
mydict = dict(dat2 = dat2, dests = dests)
dview.push(mydict)
def f(id):
sub = dat2.loc[dat2.Dest == dests[id],:]
sub = sm.add_constant(sub)
if not 'const' in sub.columns:
return None
model = sm.OLS(sub.DepDelay, sub.reindex(columns=['const','Year']))
results = model.fit()
return results.params
import time
time.time()
parallel_result = lview.map(f, range(len(dests)))
#result = map(f, range(len(dests)))
time.time()
# some NaN values because all 'Year' values are the same for some destinations
parallel_result
And we'll stop our cluster.
ipcluster stop
Let's see a basic example of doing an analysis in R across multiple cores on multiple nodes. We'll use the airline departure data in bayArea.csv.
We'll do this interactively though often this sort of thing would be done via a batch job.
# remember to do I/O off scratch
cp bayArea.csv /global/scratch/jpduncan/.
srun -A co_stat -p savio2 --nodes=2 --ntasks-per-node=24 -t 30:0 --pty bash
module load r r-packages
mpirun R CMD BATCH --no-save parallel-multi.R parallel-multi.Rout &
Now here's the R code (see parallel-multi.R) we're running:
library(doMPI)
cl = startMPIcluster() # by default will start one fewer worker, using one for the main
registerDoMPI(cl)
clusterSize(cl) # just to check
dat <- read.csv('/global/scratch/jpduncan/bayArea.csv', header = FALSE,
stringsAsFactors = FALSE)
names(dat)[16:18] <- c('delay', 'origin', 'dest')
table(dat$dest)
destVals <- unique(dat$dest)
# restrict to only columns we need to reduce copying time
dat2 <- subset(dat, select = c('delay', 'origin', 'dest'))
# some overhead in copying 'dat2' to worker processes...
results <- foreach(destVal = destVals) %dopar% {
sub <- subset(dat2, dest == destVal)
summary(sub$delay)
}
results
closeCluster(cl)
mpi.quit()
If you just want to parallelize within a node:
srun -A co_stat -p savio2 --nodes=1 -t 30:0 --pty bash
module load r
R CMD BATCH --no-save parallel-one.R parallel-one.Rout &
Now here's the R code (see parallel-one.R) we're running:
library(doParallel)
nCores <- Sys.getenv('SLURM_CPUS_ON_NODE')
registerDoParallel(nCores)
dat <- read.csv('/global/scratch/jpduncan/bayArea.csv', header = FALSE,
stringsAsFactors = FALSE)
names(dat)[16:18] <- c('delay', 'origin', 'dest')
table(dat$dest)
destVals <- unique(dat$dest)
results <- foreach(destVal = destVals) %dopar% {
sub <- subset(dat, dest == destVal)
summary(sub$delay)
}
results
You may also want to look into the future
package for R, which provides a user-friendly interface to parallelization across one or more nodes. Here's a tutorial on using future.
In addition to iPyParallel, one of the newer tools in the Python space is Dask, which provides out-of-the-box parallelization more easily without much setup or too much additional work. Dask, as a Python package, extends Numpy/Pandas syntax for arrays and dataframes that already exists and introduces native parallelization to these data structures, which speeds up analyses. Since Dask dataframes/arrays are descendants of the Pandas dataframe and Numpy array, they are compatible with any existing code and can serve as a plug-in replacement, with performance enhancements for multiple cores/nodes. It's also worth noting that Dask is useful for scaling up to large clusters like Savio but can also be useful for speeding up analyses on your local computer.
We have materials available from our spring 2019 training on using Dask on Savio and from a tutorial on using Dask.
We're also including some articles and documentation that may be helpful in getting started:
- Why Dask?
- Standard Dask Demo
- Why every Data Scientist should use Dask
- Dask Cheatsheet
- Detailed Dask overview video
- Check the Status and Announcements page:
- For technical issues and questions about using Savio:
- For questions about computing resources in general, including cloud computing:
- [email protected]
- office hours: office hours: Wed. 1:30-3:00 and Thur. 9:30-11:00 on Zoom
- For questions about data management (including HIPAA-protected data):
- [email protected]
- office hours: office hours: Wed. 1:30-3:00 and Thur. 9:30-11:00 on Zoom