Page contents:
1. NOISELAB COMPUTING RESOURCES
2. GUIDELINES FOR USAGE
2.5 GPU
3. DISK SPACE AND DATA DIRECTORIES
4. GRAPHICS
5. SOFTWARE
6. FILE MANAGEMENT
7. PRINTERS
8. USING 'SCREEN'
9. LONG MATLAB JOBS
10. MATLAB BENCHMARKING
11. MATLAB FOR PERSONAL MACHINES

1. NOISELAB COMPUTING RESOURCES

Computer CPUs RAM
Velella
 Dual Intel Xeon E5-2683 v4 2.1 GHz, 40MB cache, 16 cores/32 threads ea. (Total 64 threads available), 2 Titan GPU
528GB
Garibaldi
GPU nodes for machine learning with 8 x NVIDIA GeForce RTX 2080 cards
 384GB

The Noiselab computing clusters are listed above. Velella and Pupa are located in the UCSD network and require special accounts for access. The UCSD network can be accessed off-campus via VPN .

Both Velella and Pupa contain numerous scientific computing software packages, acoustic propagation codes, and inversion codes. Available software is listed in Section 5.

Velella storage: The command 'zfs list' will show available space

2. GUIDELINES FOR USAGE

We have no hard restrictions on CPU usage! Please be reasonable!
Some advice:
a) The load on both systems can be monitored with the top command.
b) If you have many jobs to run, execute them sequentially using scripting.
c) Run snippets of code to verify output before running the full (and potentially long) job.
d) A job should not take more than a day to run. Long jobs should be split into sequences that can be tested before continuing.
e) For long jobs write the results out with regular intervals (e.g. once an hour).

GPU for Anaconda

We recommend activating a separate environment for Tensorflow installation. As of Feb. 19, 2019, Tensorflow nightly build using CUDA 10.0 is unstable. Below is a guide to use CUDA 9.2 with stable Tensorflow. Modify versions as #updated. Install Conda, create Tensorflow-gpu environment:
(Optional: Create a directory to store your environment)

1. export PATH = /usr/local/cuda-9.2/bin:/home///bin/:$PATH
Replace “cuda-9.2” with desired cuda version, depends on tensorflow build.

2. export LD_LIBRARY_PATH = /usr/local/cuda-0.2/lib64:$LD_LIBRARY_PATH
Include 1 & 2 in .bash_profile or .bashrc. If so, restart before step 3.

3. wget https://repo.anaconda.com/archive/Anaconda3-2018.12-Linux-x86_64.sh
Set url to desired Anaconda version at (www.anaconda.com/distribution)

4. bash Anaconda3-2018.12-Linux-x86_64.sh

5. conda create —name
( is environment name, e.g. “deep learning”)
optional: condo create —n python=3.7 OR python=2.6

6. source activate

7. conda install tensorflow-gpu
(tensorflow builds are at www.tensorflow.org/install)

8. conda install

3. DISK SPACE AND DATA DIRECTORIES

All Noiselab acoustic experimental data is linked as /project/ in Velella and Pupa.
The folder /project/Noiselab with some data and past papers by the group.

You can see how much disk space is used by the command
df -h
You can see how much disk space you use in a directory and subdirectories
du -sh *

4. GRAPHICS

ExpanDrive is a file managment GUI which may provide easier interfacing between local machines and remote linux machines.

XQuartz provides a GUI environment for software launched remotely from linux terminal (e.g. Matlab)

5. SOFTWARE

The following software packages are installed on the Noiselab clusters:

SCIENTIFIC COMPUTING:
(1) Matlab R2016b
(2) Python (Anaconda, include the path '/usr/local/anaconda3/bin' in your search path to use it.)
(3) ...

TEXT EDITORS:
(1) emacs
(2) VIM
(3) ...

To query whether particular software or packages are installed, you may search for all or part of a filename with the command: 'rpm -qa | grep -i <part of name>'

If you would require a particular software package, you may request it via siohelp@ucsd.edu.

6. FILE MANAGEMENT

Files can be transferred between your local machine to the computer clusters in several ways.
One way is to use the secure copy (scp) commands:
(1) To copy a file from one machine to another via scp use: 'scp <directory/filename to copy> <destination_directory>'
...for example, to copy from your local machine to Velella use: 'scp <username>@velella.ucsd.edu:<directory/filename to copy from> <local_directory>'
(2) To copy a directory use the '-rp' (recursive, preserving other file attributes) option: 'scp -rp <directory_to_copy> <destination_directory>'

A nicer way is to use a GUI program. One program that works well is FileZilla. It provides a GUI file management interface between your local machine and ftp (sftp) servers. It works across macOS/linus/unix/windows platforms and is free and open source. To communicate with the clusters, you use SFTP (port 22).

There are also a number of intra-server (linux) GUI’s, and at least one is already installed on Velella: Nautilus, though this requires graphics rendering (xQuartz).

6. PRINTERS

The following Noiselab printers are located in Spiess Hall:

(1) Color printer: HP LaserJet 600 color MFP M775: Web address: mpl-lwwaklabd.ucsd.edu, IP Address:172.16.128.18
printer name on linux systems: lwwaklabd
(2) B/W printer: HP LaserJet Enterprise M609: mpl-lwwaklab.ucsd.edu, IP Address:172.16.128.16

7. USING 'SCREEN'

By using 'screen' you can have a termial running even after you have disconnected. When you login again you can reattach the screen. http://www.rackaid.com/resources/linux-screen-tutorial-and-how-to/

To start a new screen, simply type 'screen' at the prompt. You can start a new process in here and detach the screen by pressing Ctrl-a-d. You can logout of the main shell and screen will still be open. To see all open screens, type 'screen -ls'. If you're attached to the screen, it displayes [Attached], else it displays [Detached]. To reattach a detached screen, type ' screen -r <screen name>' You can complete screen name by tab. To exit a screen, just type 'exit' inside the screen.

8. LONG MATLAB JOBS

Before you start a long job remember to make many small tests--- It is a lot easier to find bugs in small runs!

For a long job you can save the output to mat file and then later read that into matlab for post-processing (recommended). Alternatively you you can print your figures to a file.

9. MATLAB BENCHMARKING

mean(bench(10))
LU FFT ODE Sparse 2D 3D
Pupa 0.1918 0.1383 0.1300 0.1533 3.1288 5.3459 (Aug 2017)
Velella 0.0654 0.1102 0.0656 0.1223 2.6750 4.9456 (Aug 2017)

10. MATLAB FOR PERSONAL MACHINES

The University of California San Diego has a Total Academic Headcount license for Matlab. Multiple Matlab versions, toolboxes, and licenses are
available for download for UCSD students, faculty, and staff for both univerity-owned and personal computers.

11 SAGA benchmark:

cd saga/examples/ramgeo/vertical/
saga tc1v1lay ramgeo

saga/ramgeo running tc1v1lay from the inversion workshop
(10000 forward models) on one CPU.
CPU Times:

astartes 9603.1 25 jan 01
turrid 23821.9
heart 18724 s
heart (f77 -fast ) 14812 s
heart (f77 -fast Solaris 2.8) 5297.4 (27 Dec 02)
ark (David B) f77-fast 26094 s
sundial 13046.9 25 jan
baby 2368.8 s (g77)
baby 1771 s ifc (22 Jan 03)
baby 1434.4 (ifc -O -axW -zero -w, 2 Oct 03)
baby 1475 (ifort -O -axW -u -w -posixlib -Vaxlib -CB, Feb 06)
pupa 1290.6 (ifc -O -axW -zero -w, 2 Oct 03)
occam (Mac G4 1.5GHz laptop) 4159 s (xlf -O )
G5 2 GHz Mac 1700 (xlf -O)
chenfen laptop(Mac intel ) 1237 (ifort aug 06)
Peter's laptop (Mac pro book 2.33 Ghz Core 2 duo ) 675 (ifort 9)
Peter's laptop (Mac pro book 2.33 Ghz Core 2 duo ) 637 (ifort 10.1) 27Mar 07
Mac workstaion 8GB, 3GHz 2xdual Xeon 531
Nerite 2Xquad 568 (ifort v10) 2008
Nerite 2Xquad 575 (ifort v11) June2009
Caglars laptop (Mac pro book 2.5 Ghz Core 2 duo , ifort 11) 521 June09
Peter laptop (Mac pro book 3.06 Ghz Core 2 duo , ifort 10) 434 June09
Peter laptop (Mac pro book 3.06 Ghz Core 2 duo , ifort 11) 425 July09
Peter laptop (Mac pro book 3.06 Ghz Core 2 duo , ifort 11.1 Mac10.6) 418 Oct09/ Nov10
Peter laptop (Mac pro book 3.06 Ghz Core 2 duo , ifort 12 Mac10.6) 415 Nov10
Pupa (2Xquad) ifort v11 478 Feb10
Pupa (2Xquad) ifort v12 450 Nov10
Peter Laptop (Mac book retina 2.6 Ghz Intel Cre i7; ifort 12 Mac 10.8), 303 s Aug 2012
Peter Laptop (Mac book retina 2.6 Ghz Intel Core i7; ifort 12 Mac 10.9), 271 s Nov 2013
Velella (Linux, Xeon E5-2683 v4 2.1 GHz; gfortran 7.3 ), 483 s Aug 2018