Site Tools


compute_resources

Computing Resources

The CS Dept. deploys general purpose compute and GPU compute servers, as well as specialized GPU servers, all running Linux. We also have database, NX/Nomachine (remote Linux desktop), a Windows desktop server, and Docker servers. This section describes the general use and research use servers (see other sections for database and other server information).

portal load balanced servers

The portal servers are general purpose servers, running Linux, into which anyone can login. They are available for general use. They are also the “jump off” point for off Grounds connections to the CS network. The servers are in a load balanced cluster, and are accessed through ssh to portal.cs.virginia.edu. See: The portal cluster

Use these servers to code, compile, test, etc.. However these are not meant for long running processes that will use excessive resources. Computationally expensive processes should be run on other servers listed below.

(all GPU, Memory, CPU, etc. counts are per node).

Hostname Memory (GB) CPU Type CPUs Cores/CPU Threads/Core Total Cores
portal[01-12] 132-256 Intel 1 8 2 16

GPU servers

The gpusrv* servers are general purpose servers, running Linux, that contain GPUs into which anyone can login (via 'ssh'). They are intended for code development, testing, and short computations. Long running computations are discouraged, and are better suited to one of the GPU servers controlled by the job scheduler (see the SLURM section below).

(all GPU, Memory, CPU, etc. counts are per node).

Hostname Memory (GB) CPU Type CPUs Cores/CPU Threads/Core Total Cores GPUs GPU Type GPU RAM (GB)
gpusrv[01-02,04-08] 256 Intel 1 10 2 20 4 Nvidia RTX 2080Ti 11
gpusrv03 256 Intel 1 10 2 20 3 Nvidia RTX 2080Ti 11
gpusrv[09-16] 512 Intel 2 10 2 40 4 Nvidia RTX 4000 8
gpusrv[17-18] 128 Intel 1 10 2 20 4 Nvidia RTX 2080Ti 11

Docker servers

The docker servers, running Linux, allow users to instantiate a docker container without the need for super-user (root) privileges. Once a user uses 'ssh' to login to the server, the user will execute sudo /usr/bin/docker to create a Docker container. Note that the slurm[1-5] nodes, available through the SLURM job scheduler, also allow general users to create docker containers (see below).

(all GPU, Memory, CPU, etc. counts are per node).

Hostname Memory (GB) CPU Type CPUs Cores/CPU Threads/Core Total Cores
docker01 256 Intel 2 10 2 40
docker02 64 Intel 1 4 2 8
docker03 8 Intel 1 2 2 4

Nodes controlled by the SLURM Job Scheduler

See our main article on Slurm for more information.

These servers, running Linux, are available by submitting a job through the SLURM job scheduler. They are not available for direct logins via 'ssh'. However, users can login to these servers without a job script using the 'srun -i' direct login command. See the SLURM section for more information.

Per CS Dept. policy, servers are placed into the SLURM job scheduling queues and are available for general use. Also, per that policy, if a user in a research group that originally purchased the hardware requires exclusive use of that hardware, they can be given a reservation for that exclusive use for a specified time. Otherwise, the systems are open for use by anyone with a CS account. This policy was approved by the CS Dept. Computing Committee comprised of CS Faculty.

This policy allows servers to be used when the project group is not using them. So instead of sitting idle and consuming power and cooling, other Dept. users can benefit from the use of these systems.

All GPU, Memory, CPU, etc. counts are per node.. This list may not be up to date. Use the 'sinfo' command for current server inventory.

Hostname Mem (GB) CPU #CPUs Cores Threads / Core Total Threads GPUs GPU Type GPU RAM (GB)
adriatic[01-06] 1024 Intel 2 8 2 32 4 Nvidia RTX 4000 8
affogato[01-10] 128 Intel 2 8 2 32 0
affogato[11-15] 128 Intel 2 8 2 32 4 Nvidia GTX1080Ti 11
ai[01-04,06] 64 Intel 2 8 2 32 4 Nvidia GTX1080Ti 11
ai05 64 Intel 2 8 2 32 4 Nvidia GTX1080 11
ai[07-09] 128 Intel 2 8 2 32 4 Nvidia GTX1080Ti 11
ai10 128 Intel 2 8 2 32 4 Nvidia GTX1080 11
cheetah01 256 AMD 2 8 2 32 4 Nvidia A100 40
cheetah[02-03] 1024(2) Intel 2 18 2 72 2 Nvidia RTX 2080Ti 11
cheetah04 1024 AMD 2 64 2 256 4 Nvidia A100 80(5)
cheetah05 768 AMD 2 16 2 64 4 Nvidia A100 80(5)
cortado[01-10] 512 Intel 2 12 2 48 0
doppio[01-05] 128 Intel 2 16 2 64 0
epona 64 Intel 1 4 2 8 0
heartpiece 160 Intel 2 10 2 40 0
hydro 256 Intel 2 16 2 64 0
jaguar01 1024 Intel 2 16 2 64 4 Nvidia A40 48 (4)
jaguar02 1024 Intel 2 8 2 32 8 Nvidia A16 16
jaguar03 256 AMD 2 56 2 224 8 Nvidia RTX A4500 20
jaguar04 256 Intel 2 16 2 64 4 Nvidia A40 48 (4)
jaguar05 256 Intel 1 8 2 16 4 Quadro RTX 4000 8
jaguar06 128 Intel 2 24 2 48 2 Nvidia A40 48
jinx[01-02] 160 Intel 1 24 2 48 1 Nvidia GTX1080 11
lotus 256 Intel 2 20 2 80 8 Nvidia RTX 6000 24
lynx01 64 Intel 4 8 2 32 4 Nvidia Titan X 12
lynx[01-04] 64 Intel 4 8 2 32 4 Nvidia GTX1080Ti 11
lynx[05-07] 64 Intel 4 8 2 32 4 Nvidia P100 16
lynx[08-09] 64 Intel 4 8 2 32 0
lynx10 64 Intel 4 8 2 32 1 Nvidia GTX1080Ti 11
2 Nvidia GTX1080 8
lynx11 64 Intel 4 8 2 32 3 Nvidia Titan X 12
1 Nvidia GTX1080 8
lynx12 64 Intel 4 8 2 32 4 Nvidia Titan X 12
optane01 1024(1) Intel 2 16 2 64 0
panther01 512 Intel 1 8 2 16 0
pegasusboots 192 Intel 2 10 2 40 0
puma02 256 AMD 2 64 2 256 6 NVIDIA RTX A6000
ristretto[01-04] 128 Intel 2 6 1 12 8 Nvidia GTX1080Ti 11
serval01 2048 Intel 2 56 2 224 4 Nvidia H100 80 (5)
sds[01-02] 512 Intel 2 10 2 40 4 Nvidia RTX A4000 16
slurm[1-5](3) 512 Intel 2 12 2 24 0
titanx[01-03] 256 Intel 1 8 2 16 1 Nvidia Titan X 12
titanx[04-05] 64 Intel 1 6 2 12 1 Nvidia Titan X 12

(1) 512GB Intel Optane memory, 512 DDR4 memory

(2) In addition to 1TB of DDR4 RAM, these servers also house a 900GB Optane NVMe SSD and a 1.6TB NVMe regular SSD drive

(3) All users have 'sudo' privileges to execute /usr/bin/docker on these nodes

(4) Scaled to 96GB by pairing GPUs with NVLink technology that provides a maximum bi-directional bandwidth of 112GB/s between the paired A40s

(5) Scaled to 320GB by aggregating all four GPUs with NVLink technology

Job Scheduler Queues

See our main article on Slurm for more information on using queues (“partitions”)

The main partition contains nodes without GPUs, the gpu partition contains nodes with gpus, the nolom partition has nodes with no gpus and no time limit on jobs, and the gnolim partition has nodes with GPUs but no time limit on jobs. The time limit on jobs in the main and gpu partitions is shown using the 'sinfo' command.

Queue Nodes
main cortado[01-10], hydro, optane01, slurm[1-5]
gpu adriatic[01-06], affogato[11-15], ai[01-08], cheetah[01-05], jaguar[01-03], lotus, lynx[01-12], ristretto[01-04], sds[01-02]
nolim doppio[01-05], epona, heartpiece, pegasusboots
gnolim ai[01-10], titanx[01-05]

Group specific computing resources

Several groups in CS deploy servers that are used exclusively by that group. Approximately 40 servers are deployed in this fashion, ranging from traditional CPU servers to specialized servers containing GPU accelerators.

Docker Server

See our main article on Docker for more information.

The Docker server on hostname 'docker01' is currently available to all department members. Docker containers can also be created on nodes slurm[1-5].

compute_resources.txt · Last modified: 2024/02/20 20:21 by 127.0.0.1