Data storage¶

Video

Watch this in our courses: 2022 February, 2021 January

These days, computing is as much (or more) about data than the actual computing power. And data is more than number of petabytes: it is so easy to get it unorganized, or stored in such a way that it slows down the computation.

In this tutorial, we go over places to store data on Triton and how to choose between them. The next tutorial tells how to access it remotely.

Abstract

See the Triton quick reference
We are a standard Linux filesystem
- $HOME = /home/$USER: 10GB, backed up, not made larger
- Scratch is large but not backed up:
  - $WRKDIR = /scratch/work/$USER: Personal work directory
  - /scratch/DEPARTMENT/NAME/: Group-based shared directories (recommended for most work, group leaders can request them)
- /tmp: temporary directory, pre-user mounted in jobs and automatically cleaned up.
- /l/: local persistent storage on some group servers
- $XDG_RUNTIME_DIR: ramfs on login node
See Remote access to data for how to transfer and access the data from other computers.

Basics¶

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 ls -l or stat.

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. Check out wikipedia:Memory Hierarchy and wikipedia:List of interface bit rates.

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¶

Each storage location has different sizes, speed, types of backups, and availability. You need to balance between these.

Name	Path	Quota	Backup	Locality	Purpose
Home	`$HOME` or `/home/USERNAME/`	hard quota 10GB	Nightly	all nodes	Small user specific files, no calculation data.
Work	`$WRKDIR` or `/scratch/work/USERNAME/`	200GB and 1 million files	x	all nodes	Personal working space for every user. Calculation data etc. Quota can be increased on request.
Scratch	`/scratch/DEPT/PROJECT/`	on request	x	all nodes	Department/group specific project directories.
Local temp	`/tmp/`	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.
Local persistent	`/l/`	varies	x	dedicated group servers only	Local disk persistent storage. On servers purchased for a specific group. Not backed up.
ramfs (login nodes only)	`$XDG_RUNTIME_DIR`	limited by memory	x	single-node	Ramfs on the login node only, in-memory filesystem

Home directories¶

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 change to $WRKDIR or a group scratch directory when you first log in and start doing work. (note: home and work may be deleted six months after your account expires: use a group-based space instead).

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.

See Storage: Lustre (scratch)

Local disks¶

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 --gres=spindle.

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.

Other Aalto data storage locations¶

Aalto has other non-Triton data storage locations available. See Data storage and Data: outline, requesting space, requesting access for more info.

Quotas¶

All directories under /scratch (as well as /home) have quotas. Two quotas are set per-filesystem: disk space and file number. Quotas exist not because we need to limit space, but because we need to make people think before using large amounts of space. Ask us if you need more.

Disk quota and current usage are printed with the command quota. ‘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 member.

$ 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.

Remote access¶

The next tutorial, Remote access to data, covers accessing the data from your own computer.

Exercises¶

Most of these exercises will be specific to your local site. Use this time to review your local guides to see how they are adapted to your site.

Data storage locations:

Storage-1: Review data storage locations

(Optional) Look at the list of data storage locations above. Also look at the Data storage. Which do you think are suitable for your work? Do you need to share with others?

Storage-2: Your group’s data storage locations

Ask your group what they use and if you can use that, too.

Misc:

Storage-3: Common errors

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 -X doesn’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.

Solution

All of these can be caused by exceeding the quota.

(don’t worry, “can’t log in” doesn’t apply to basic ssh login, so you can always still fix it yourself)

About filesystem performance:

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.

Storage-4: strace and I/O operations

Use strace -c to compare the number of system calls in ls, ls -l, ls --no-color, and ls --color on a directory with many files. On Triton, 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?

Storage-5: strace and time

Using strace -c, compare the times of find and lfs find on the directory mentioned above. Why is it different?

(advanced) Storage-6: Benchmarking

(this exercise requires slurm knowledge from future tutorials and also other slurm knowledge).

Clone the https://github.com/AaltoSciComp/hpc-examples/ git repository to your personal work directory. Change to the io directory. Create a temporary directory and…

Run create_iodata.sh to make some data files in data/
Compare the IO operations of find and lfs find on this directory.
use the iotest.sh script 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.
Use tar to compress the data while it is on lustre. Unpack this tar archive to local storage, do the operations, then remove. Compare to previous strategies.