Narval
| Availability: October, 2021 |
| Login node: To be announced |
| Globus endpoint: To be announced |
| Data transfer node (rsync, scp, sftp,...): To be announced |
Narval is a general purpose cluster designed for a variety of workloads; it is located at the École de technologie supérieure in Montreal. The cluster is named in honour of the narwhal, a species of whale which has occasionally been observed in the Gulf of St. Lawrence.
Site-specific policies
By policy, Narval's compute nodes cannot access the internet. If you need an exception to this rule, contact technical support with information about the IP address, port number(s) and protocol(s) needed as well as the duration and a contact person.
Crontab is not offered on Narval.
Each job on Narval should have a duration of at least one hour (five minutes for test jobs) and you cannot have more than 1000 jobs, running and queued, at any given moment. The maximum duration for a job on Narval is 7 days (168 hours).
Storage
| HOME Lustre filesystem, ~100 TB of space |
|
| SCRATCH Lustre filesystem, ~5 PB of space |
|
| PROJECT Lustre filesystem, ~15 PB of space |
|
For transferring data via Globus, you should use the endpoint specified at the top of this page, while for tools like rsync and scp you can use a login node.
High-performance interconnect
The InfiniBand Mellanox HDR network links together all of the nodes of the cluster. Each hub of 40 HDR ports (200 Gb/s) can connect up to 66 nodes with HDR100 (100 Gb/s) with 33 HDR links divided in two (2) by special cables. The seven (7) remaining HDR links allow the hub to be connected to a rack containing the seven (7) central HDR InfiniBand hubs. The islands of nodes are therefore connected by a maximum blocking factor of 33:7 (4.7:1). In contrast, the storage servers are connected by a much lower blocking factor in order to maximize the performance.
In practice the Narval racks contain islands of 48 or 56 regular CPU nodes. It is therefore possible to run parallel jobs using up to 3584 cores with a non-blocking network. For larger jobs or ones which are distributed in a fragmented manner across the network, the blocking factor is 4.7:1. The inter-connect remains a high-performance one nonetheless.
Node Characteristics
| nodes | cores | available memory | CPU | storage | GPU |
|---|---|---|---|---|---|
| 1109 | 64 | ~256G | 2 x AMD Rome 7532 @ 2.40 GHz 256M cache L3 | 1 x 960G SSD | - |
| 33 | ~2048G | ||||
| 158 | 48 | ~512G | 2 x AMD Milan 7413 @ 2.65 GHz 128M cache L3 | 1 x SSD of 3.84 TB | 4 x NVidia A100 (40 GB memory) |
AMD processors
Supported instructions sets
Narval is equipped with 2nd and 3rd generation AMD EPYC processors which support AVX2 instructions. This instruction set is identical to the one used by INTEL processors on the Béluga, Cedar, Graham and Niagara nodes. The Narval AMD processors and the Intel processors of the Haswell generation are both restricted to AVX2 and earlier instructions. As such, applications compiled on a Cedar or Graham Broadwell node should function on Narval, otherwise you will need to recompile with the appropriate architecture parameters (see Intel compilers below).
Unfortunately, Narval's AMD processors are incompatible with the AVX512 instruction set (from https://en.wikipedia.org/wiki/Skylake Skylake] onwards] on the Béluga and Niagara nodes, as well as the more recent nodes on Cedar and Graham. Consequently, applications compiled on those relatively recent Intel nodes will probably not function on Narval and recompiling will certainly be necessary (see Intel compilers below).
Intel compilers
The Intel compilers can very well compile applications for Narval's AMD processors with the less recent AVX2 instruction sets. To do so, use the -march=core-avx2 option to produce executables which are compatible with both Intel and AMD processors.
However, the applications will not operate on Narval if your code was compiled on a system using one or more -xXXXX options such as -xCORE-AVX2 because the Intel compilers add extra instructions to verify that the processor is by Intel. This beinig said, the -xHOST on Narval becomes the same as -march=core-avx2.
Software environments
StdEnv/2020 is the standard software environment on Narval; previous versions have been blocked intentionally. If you need an application only available with an older standard environment, please write to Technical support.
BLAS and LAPACK libraries
The Intel MKL library works with AMD processors, although not in an optimal way. We are presently examining the option of using the FlexiBLAS library when installing the next software versions; this would automatically load the appropriate library, depending on the processor's provider. If you think that your code is espacially sensitive to the performance of BLAS or LAPACK, please report this to Technical support.