Page MenuHomeSoftware Heritage

prepare quote for "granet2", next gen swh-graph compression server
Closed, MigratedEdits Locked

Description

The general idea is to split the workload between:

  • granet1 (current server): will be used to serve the compressed graph in production
  • granet2 (to be provisioned): will be used to periodically export/compress the graph from the archive, and sporadic R&D experiments

that way we can play with the graph compression stack without impacting the planned in-production use of swh-graph (T2220).

Hardware specs (WIP)

  • RAM: 2-3 TiB
    • 2 TiB is a comfortable estimate for being able to compress the graph for now and in the foreseeable future
    • 3 TiB is what we want to be able to extend to in the future, ideally without having to swap RAM (this impacts both the kind of DIMMs we want and possibly CPUs)
  • CPUs: 48 cores (same as granet1) should be plenty
  • Disks:
    • SSD: 20 TB
    • HDD: 50 TB

      General comments on storage:
      • For the intended use of granet2 (!= granet1), all work space on disk could be considered scratch space with no redundancy, if needed to keep storage costs under control
      • If the new operation requirements for granet1 allows that, we should consider moving disks (of any kind) from granet1 to granet2
      • In hindsight, buying "read-intensive" for granet1 might have been a mistake. So we should avoid that this time

Related Objects

Event Timeline

zack triaged this task as High priority.Oct 8 2021, 4:04 PM
zack created this task.

I've tried to generate a server config, out of the HPE ProLiant DL380 G10 (Intel based) and DL385 G10+ (AMD EPYC based) configs available on MATINFO5.

However, I've found the following issues:

  • Intel-based config: impossible to select CPUs able to address more than 1 TB of memory ("gen 3" processors), compromising the extensibility of the system to 3 TB RAM (max out at 2 TB).
  • Both configs: impossible to add the SFF (2.5") drive cage in the 2 CPU configs (unable to select the "balanced" NVMe backplane) to mix LFF (3.5") and SFF drives.
  • Both configs: unable to add full SFF drive cages to attempt a full-NVMe config.

I've contacted the HPE representatives about these issues, and asked them for the following configs based around a dual AMD EPYC system:

  • 2TB mem, 4*6.4TB NVMe SFF, 6*12TB SAS LFF
  • 2TB mem, 12*6.4 TB NVMe SFF
  • 3TB mem, 4*6.4TB NVMe SFF, 6*12TB SAS LFF
  • 3TB mem, 12*6.4 TB NVMe SFF