You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 19 Next »

Questions/Suggestions

hpc@awi.de

Hardware

  • 2x interactive login nodes

    • 64bit-processor, AMD Rome (Epyc 7702)

    • 2 sockets per node
    • 64 core, cTDP reduced from 200W to 165W
    • → 128 cores/node
    • 2GHz (3.3GHz Boost)
    • 512 GiB RAM (8 DIMMs per socket with 32GiB DDR4-3200)
    • 2x SSD 480 GB SATA
    • 1x SSD 1,92 TB
  • 1x interactive GPU node
    • 2x NVDIA A40
    • 2x AMD Rome (Epyc 7702) 
    • 512 GiB RAM (8 DIMMs per socket with 32GiB DDR4-3200)
    • for testing, Jupyter notebooks, Matlab, ...

  • 240x standard compute nodes (je vier in einem NEC HPC2824Ri-2)

    • 256 GiB RAM (8 DIMMs  per socket with 16GiB DDR4-3200)

    • 512 GiB NVMe pro Knoten

  • 4x fat nodes (jeweils in einem NEC HPC2104Ri-1)
    • 4 TiB RAM (16 DIMMs per socket with 128 GiB DDR4-3200)
    • 512 GiB NVMe
    • 7.68 TiB NVMe
  • 1x GPU nodes (1x HPC22G8Ri)
    • 4x Nvidia A100/80
    • 2x AMD Rome (Epyc 7702) 
    • 512/1024 GiB RAM
    • 3x3.84 TiB = 11.52 TiB NVMe
    • More GPU nodes will follow later, after we gained first experience of what you really need, and to offer most recent hardware
  • 3x Management nodes
    •  one AMD 7302P (NEC HPC2104Ri-1) each
    • 1 socket
    • 16 cores
    • 128 GiB RAM (8 DIMMs  per socket with 16GiB DDR4-3200)
    • 2x SATA 960 GiB SSD
    • 1x 1.92 TiB NVMe
  • File Storage:
    • IBM Spectrum Scale (GxFS)
    • Tier 1: 220 TiB NVMe as fast cache and/or burst buffer
    • Tier 2: ~5.38 PiB NL-SAS HDD (NetApp EF300)
    • future extension of both tiers (capacity, bandwidth) is possible
  • Fast interconnect: HDR Infiniband (100 Gb)
  • All nodes connected to /isibhv (NFS, 10 Gb ethernet)
  • Alma Linux ("free RedHat", version 8.x)

The FESOM2 Benchmark we used for the procurement on 240 Albedo nodes compares to 800 Ollie nodes.

Preliminary schedule for the transition Ollie → Albedo

Disclaimer: We cannot guarantee the following time frame! It is an optimistic view and can be disturbed by further lock downs in China, a ship may block the Suez channel, or it can be just small technical issues that prevent Albedo from smooth operation even with all hardware at AWI in time . Any step can be delayed by weeks or months! Except the first "Now" step, which is in your hands ;-)

  • Now Please start to prepare your directories on /work/ollie for transfer to Albedo!
    • Delete data you do not need to keep,
    • copy tarballs of valuable data to the tape archive (you should do this anyhow, /work has no backup),
    • place a copy of data which you need continuously with fast access to /isibhv.
    • We will not transfer data automatically from Ollie to Albedo!! This is a chance to clean up ;-)
  • until April  NEC investigates if OmniPath is an alternative for Mellanox Infinband as the fast network in Albedo, because Infiniband would be delivered with a large delay
  • May
    Albedo installed at AWI  with "slow" Gigabit  (10 Gb/s) network and, if recommended by NEC, fast OmniPath (100 Gb/s)
  • June / July
    Albedo open for power users
    Start to copy data from Ollie to Albedo (Note: must be done by yourself, there is no automatic data migration!)
    The more Albedo nodes are powered on, the more Ollie nodes have to be switched off. We have no schedule for this, but would decide depending on how useful Albedo already is for how many users. Of course, Albedo has a far better ratio of computing vs. electric power.
  • August
    Albedo open for all users
  • 30. June
    Ollie hardware support ends. As a prolongation would cost approx. 200.000€, we keep Ollie running w/o support, but with the option to have basic components (file system, network) repaired in case of an emergency (maintenance on demand). So far, we have very good experiences with the service of MegWare and are confident that we found a solution that is cost efficient and sustainable.
  • August Downtime to add Infiniband

More details

Basically, all hardware is in place at NEC's technical center in Germany except the CPUs (expected end of March) and Mellanox Infiniband cables (expected end of July or even later). Concerning the CPUs, NEC is confident that the date of delivery is met, however, there can always be a bad surprise. For Mellanox, the situation is worse. As Mellanox now belongs to NVIDIA, all customers for NVIDIA's flagship DGX are served first. Therefor, NEC is currently investigating if OmniPath offers similar performance for our benchmarks and for the parallel filesystem GxFS (5PB work/scratch). Depending on the outcome, we would start with OmniPath in April, or with just Gigabit network and with a downtime in August to add Infiniband.

Mellanox Infiniband or Cornelis Networks OmniPath? You know OmniPath from Ollie. We have mixed experiences. The MPI performance is very good for the typical codes at AWI, however, Intel declared it deprecated and as a consequence, bugs in BeeGFS in the interplay with OmniPath were never fixed. At the time of the procurement, Infiniband was the only option. However, as NVIDIA acquired Mellanox, Intel decided to revive OmniPath with a spin-off, Cornelis Networks. Due to NVIDIA's policy and the long delays to obtain the infiniband hardware, OmniPath with similar performance seems to have a good new start and is already part of new installations. NEC is careful and they are testing OmniPath thoroughly in new hardware at their test center before they give a recommendation for Albedo. Differences matter: OmniPath puts more load onto the CPUs of the compute nodes and file servers, but comes with better latencies in return. Also, internal configuration options differ.   

  • No labels