In this tutorial, we go over places to store data on Triton and how to access it remotely.
Optimizing data storage isn’t very glamorous, but is an important part of high-performance computing.
Triton has various ways to store data. Each has a purpose, and when you are dealing with large data sets or intensive I/O, efficiency becomes important.
Roughly, we have small home directories (only for configuration files), large Lustre (scratch and work, large, primary calculation data), and special places for scratch during computations (local disks). At Aalto, there is aalto home, project, and archive directories which, unlike Triton, are backed up but don’t scale to the size of Triton.
A file consists of its contents and metadata. The metadata is information
like user, group, timestamps, permissions. To view metadata, use
Filesystem performance can be measured by both IOPS (input-output
operations per second) and stream I/O speed.
/usr/bin/time -v can
give you some hints here. You can see the profiling page for more information.
Think about I/O before you start! - General notes¶
When people think of computer speed, they usually think of CPU speed. But this is missing an important factor: How fast can data get to the CPU? In many cases, input/output (IO) is the true bottleneck and must be considered just as much as processor speed. In fact, modern computers and especially GPUs are so fast that it becomes very easy for a few GPUs with bad data access patterns to bring the cluster down for everyone.
The solution is similar to how you have to consider memory: There are different types of filesystems with different tradeoffs between speed, size, and performance, and you have to use the right one for the right job. Often times. So you have to use several in tandem: For example, store original data on archive, put your working copy on scratch, and maybe even make a per-calculation copy on local disks.
The following factors are useful to consider:
- How much I/O are you doing in the first place? Do you continually re-read the same data?
- What’s the pattern of your I/O and which filesystem is best for it? If you read all at once, scratch is fine. But if there are many small files or random access, local disks may help.
- Do you write log files/checkpoints more often than is needed?
- Some programs use local disk as swap-space. Only turn on if you know it is reasonable.
There’s a checklist in the storage details page.
Avoid many small files! Use a few big ones instead. (we have a dedicated page on the matter)
Available data storage options¶
||hard quota 10GB||Nightly||all nodes||Small user specific files, no calculation data.|
||200GB and 1 million files||x||all nodes||Personal working space for every user. Calculation data etc. Quota can be increased on request.|
||on request||x||all nodes||Department/group specific project directories.|
||limited by disk size||x||single-node||Primary (and usually fastest) place for single-node calculation data. Removed once user’s jobs are finished on the node.|
||varies||x||dedicated group servers only||Local disk persistent storage. On servers purchased for a specific group. Not backed up.|
|ramfs (login nodes only)||
||limited by memory||x||single-node||Ramfs on the login node only, in-memory filesystem|
The place you start when you log in. Home directory should be used for init files, small config files, etc. It is however not suitable for storing calculation data. Home directories are backed up daily. You usually want to use scratch instead.
scratch and work: Lustre¶
Scratch is the big, high-performance, 2PB Triton storage. It is the primary
place for calculations, data analyzes etc. It is not backed up but is
reliable against hardware failures (RAID6, redundant servers), but
not safe against human error.. It is
shared on all nodes, and has very fast access. It is divided into two
parts, scratch (by groups) and work (per-user). In general, always
$WRKDIR or a group
scratch directory when you first
log in and start doing work.
Lustre separates metadata and contents onto separate object and metadata servers. This allows fast access to large files, but induces a larger overhead than normal filesystems. See our small files page for more information.
Local disks are on each node separately. It is used for the fastest I/Os
with single-node jobs and is cleaned up after job is finished. Since 2019,
things have gotten a bit more complicated given that our newest (skl) nodes
don’t have local disks. If you want to ensure you have local storage,
submit your job with
See the Compute node local drives page for further details and script examples.
ramfs - fast and highly temporary storage¶
On login nodes only,
$XDG_RUNTIME_DIR is a ramfs, which means that it looks like files
but is stored only in memory. Because of this, it is extremely fast,
but has no persistence whatsoever. Use it if you have to make small
temporary files that don’t need to last long. Note that this is no
different than just holding the data in memory, if you can hold in
memory that’s better.
All directories under
/scratch (as well as
/home) have quotas. Two
quotas are set per-filesystem: disk space and file number.
Disk quota and current usage are printed with the command
‘space’ is for the disk space and ‘files’ for the total number of files
limit. There is a separate quota for groups on which the user is a
$ quota User quotas for darstr1 Filesystem space quota limit grace files quota limit grace /home 484M 977M 1075M 10264 0 0 /scratch 3237G 200G 210G - 158M 1M 1M - Group quotas Filesystem group space quota limit grace files quota limit grace /scratch domain users 132G 10M 10M - 310M 5000 5000 - /scratch some-group 534G 524G 524G - 7534 1000M 1000M - /scratch other-group 16T 20T 20T - 1088M 5M 5M -
If you get a quota error, see the quotas page for a solution.
Accessing and transferring files remotely¶
Transferring files to/from Triton is exactly the same as any other remote Linux server.
Remote mounting using SMB¶
By far, remote mounting of files is the easiest method to transfer files. If you are
not on the Aalto networks (wired,
Aalto-managed laptop), connect to the Aalto VPN first. Note that
this is automatically done on some department workstations (see
below) - if not, request it!
The scratch filesystem can be remote mounted using SMB inside secure Aalto networks at the URLs
On different operating systems:
- Linux (Ubuntu for example): File manager (Nautilus) → File →
Connect to server. Use the
- Windows: In the file manager, go to Computer (in menu bar on top, at
least in Windows 10) → Map Network Drive) and “Map Network Drive”.
In Windows 10 → “This PC” → right click → “Add Network Location”.
(Note that this is different from right-click “Add network location”
which just makes a folder link and has had some problems in the past.)
Use the URLs above but replace
\. For example,
- Mac: Finder → Go → Connect to Server. Use the
Depending on your OS, you may need to use either your username
Remote mounting using sshfs¶
sshfs is a neat program that lets you mount remote filesystems via
ssh only. It is well-supported in Linux, and somewhat on other
operating systems. It’s true advantage is that you can mount any
remote ssh server - it doesn’t have to be set up specially for SMB or
any other type of mounting. On Ubuntu, you can mount by “File → Connect to
server” and using
The below uses command line programs to do the same, and makes the
triton_work on your local computer access all files in
/scratch/work/USERNAME. Can be done with other folders.:
mkdir triton_work sshfs USERNAME@triton.aalto.fi:/scratch/work/USERNAME triton_work
ssh binds together many ways of accessing Triton, with a
similar syntax and options.
ssh is a very important program and
binds together all types of remote access, and learning to use it well
will help you for a long time.
The SFTP protocol uses ssh to transfer files. On Linux and Mac, the
sftp command line program are the must fundamental way to do this,
and are available everywhere.
Below is an example of the “raw” SFTP usage:
# Copying from HOME to local PC [email protected] $ sftp [email protected]:filename Connected to triton.aalto.fi. Fetching /home/user12/filename to filename # copying to HOME [email protected] $ sftp -b - [email protected] <<< 'put testCluster.m' sftp> put foo # copying to WRKDIR [email protected] $ sftp -b - [email protected]:/scratch/work/USERNAME/ <<< 'put testCluster.m' ...
With all modern OS it is also possible to just open your OS file manager (e.g. Nautilus on Linux) and just put as address in the bar:
If you are connecting from remote and cannot use the VPN, you can connect instead to department machines like kosh.aalto.fi, taltta.aalto.fi, amor.org.aalto.fi (for NBE). The port is 22. Note: If you do not see your shared folder, you need to manually specify the full path (i.e. the folder is there, just not yet visible).
Rsync is similar to sftp, but is smarter at restarting files. Use rsync
for large file transfers.
rsync actually uses the ssh protocol so
rsync from anywhere you can
rsync is installed
by default on Linux and Mac terminals. On Windows machines we recommend using GIT-bash.
While there are better places on the internet to read about rsync, it is good to try it out to sychronise a local folder on your triton’s scratch. Sometimes the issue with copying files is related to group permissions. This command takes care of permissions and makes sure that all your local files are identical (= same MD5 fingerprint) to your remote files:
rsync -avzc -e "ssh" --chmod=g+s,g+rw --group=GROUPNAME PATHTOLOCALFOLDER USERNAME@triton.aalto.fi:/scratch/DEPT/PROJECTNAME/REMOTEFOLDER/
Replace the bits in CAPS with your own case. Briefly,
-a tries to preserve all attributes of the file,
-v increases verbosity to see what rsync is doing,
-z uses compression,
-c skips files that have identical MD5 checksum,
-e specifies to use ssh (not necessary but needed for the commands coming after),
--chmod sets the group permissions to shared (as common practice on scratch project folders), and
--group sets the groupname to the group you belong to (note that GROUPNAME == PROJECTNAME on our scratch filesystem).
If you want to just check that your local files are different from the remote ones, you can run rsync in “dry run” so that you only see what the command would do, without actually doing anything.:
rsync --dry-run -avzc ...
Sometimes you want to copy only certain files. E.g. go through all folders, consider only files ending with
rsync -avzc --include '*/' --include '*.py' --exclude '*' ...
Sometimes you want to copy only files under a certain size (e.g. 100MB):
rsync -avzc --max-size=100m ...
Rsync does NOT delete files by default, i.e. if you delete a file from the local folder, the remote file will not be deleted automatically, unless you specify the
Please note that when working with files containing code or simple text, git is a better option to synchronise your local folder with your remote one, because not only it will keep the two folders in sycn, but you will also gain version controlling so that you can revert to previous version of your code, or txt/csv files.
Accessing files from Department workstations¶
This varies per department, with some strategies that work from everywhere.
These mounts that are already on workstations require a valid Kerberos
ticket (usually generated when you log in). On long sessions these might
expire, and you have to renew them with
kinit to keep going.
The staff shell server
taltta.aalto.fi has scratch and work mounted
/m/triton, and department directories are also in the standard
Work directories are available at
/m/nbe/work and group scratch
Directories available on demand through SSHFS. See the Data transferring page at PHYS Intranet (accessible by PHYS users only).
Work directories are available at
/m/cs/work/, and group scratch
strace is a command which tracks system calls, basically the
number of times the operating system has to do something. It can be
used as a rudimentary way to see how much I/O load there is.
strace -cto compare the number of system calls in
ls --no-color, and
ls --color. You can use the directory
/scratch/scip/lustre_2017/many-files/as a place with many files in it. How many system calls per file were there for each option?
strace -c, compare the times of
lfs findon the directory mentioned above. Why is it different?
- (Advanced, requires slurm knowledge from future tutorials) You
will find some sample files in
/scratch/scip/hpc-examples/io. Create a temporary directory and…
create_iodata.shto make some data files in
- Compare the IO operations of
lfs findon this directory.
- use the
iotest.shscript to do some basic analysis. How long does it take? Submit it as a slurm batch job.
- Modify the iotest.sh script to copy the
data/directory to local storage, do the operations, then remove the data. Compare to previous strategy.
tarto compress the data while it is on lustre. Unpack this tar archive to local storage, do the operations, then remove. Compare to previous strategies.
- Mount your work directory by SMB - and alternatively sftp or sshfs - and transfer a file to Triton.
Note that you must be on
aaltowith Aalto laptop, or connected to the Aalto VPN.
- (Advanced) If you have a Linux on Mac computer, study the
rsyncmanual page and try to transfer a file.
- What do all of the following have in common?
- A job is submitted but fails with no output or messages.
- I can’t start a Jupyter server on jupyter.triton.
- Some files are randomly empty. Or the file had content, I tried to save it again, and now it’s empty!
- I can’t log in.
- I can log in with ssh, but
ssh -Xdoesn’t work for graphical programs.
- I get an error message about corruption, such as
InvalidArchiveError("Error with archive ... You probably need to delete and re-download or re-create this file.
- I can’t install my own Python/R/etc libraries.