Project

General

Profile

Actions

Task #8069

closed

Investigate potential bottleneck on storage/CEPH at DCL

Added by Timothée Floure almost 4 years ago. Updated 3 months ago.

Status:
Closed
Priority:
Normal
Assignee:
-
Target version:
-
Start date:
05/27/2020
Due date:
% Done:

0%

Estimated time:
PM Check date:
Actions #1

Updated by Timothée Floure almost 4 years ago

Our hardware:

Actions #2

Updated by Nico Schottelius almost 4 years ago

Some questions we should be able to answer:

Real scenarios

NOTE: assuming all disks running at 'full speed'.
NOTE: big big unknown here is how the cache of the RAID controller behave.
NOTE: unknown IOPS limitations on raid controllers. <--- TODO, more important than bandwith!

  • The R710 server has 8 disks slots (supposedly with a h700 controller). Given that we fully populate the server, what is the maximum bandwidth available per OSD running on that machine?
    -> 4 GB/s from PCIe but 3GB/s for SATA -> 375 MB/s per disk modulo caching from RAID controller.
  • The R815 has 6 disk slots (is that true? -> Balazs). Same question as above.
    -> SAS 6GB/s but 3GB/s for SATA -> 500 MB/s per disk, module caching from RAID controller.
  • What about an R815 with an md array (12x 3.5" HDD via SAS cable attached to H800)
    -> 4 GB/s from PCIe connector (SAS supports 6GB/s) -> 333 MB/s per device.
    • Is the bottleneck likely a) the disk b) the controller c) the network of the server d) another component in the server
      - 10 Gbps = 1.25 GB/s = 104 MB/s per disk at full speed.
      - Controller PCIe limits at 500 MB/s per disk at full speed.
      -> Bottleneck likely to be on disk or network.
  • Given an Arista 7050 and an imaginary bandwidth per disk of 50 MB/s, how many disks can we run on one 7050?
    - The Arista is supposed to handle 1.04 Tbps = 130000 MB/s = 2600 * 50 MB/s => not an issue.
  • Is the PCI-E bus (it's not a bus anymore - afair it's point-to-point) on either server model a limitation?
    • It provides access to networking, disks and has an interconnect to the cpus

-> No worried, but TODO.

  • We are using ceph bluestore (https://ceph.io/community/new-luminous-bluestore/)
    • Does it make sense to switch our storage model to use 2 SSDs (f.i. 1TB) in a raid1 in front of HDDs and drop the distinction of HDD/SSD?
      • raid1 is needed as on the failure of the SSD all osds that have the rocksdb/bluefs on it fail

-> TODO

(skip answers if they are too far from what you can gather)

Actions #3

Updated by Timothée Floure almost 4 years ago

Regarding the RAID controllers:

  • RAID0 (striping - redundancy is handled by CEPH across physical servers).
  • Some controllers are battery-backed:
    - Likely write-back cache.
  • Some are not:
    - Likely write-though cache.
    - .. or forced WB via BIOS/firmware setting?
  • Read cache defaults to 'Adaptive Read Ahead': When selected, the controller begins using Read-Ahead if the two most recent disk accesses occurred in sequential sectors.
    - Fairly useless for random reads.
Actions #4

Updated by Timothée Floure almost 4 years ago

Regarding PCIe AND SAS/SATA:

  • Controllers are connected on x8 PCIe 2.0 => 500 MB/s per-lane for PCIe 2.0 -> x8 = 4 GB/s
  • 6 GB/s SAS 2.0 connectivity -> how is this split between disks? Should be fine anyway.
    - perc h700 supports SATA 3GB/s, perch800 does not support SATA.
  • How are our network cards connected? Should be fine anyway. 10Gbpe = 1.25 GB/s -> even PCIe 2.0 4x is more than enough.
Actions #5

Updated by Timothée Floure almost 4 years ago

I'll be AFK for a little while: the big pain point is the hardware RAID controller.

  • Unknown effect on IOPS (needs more digging, not obvious).
    • The internet says (reddit, random wikis, CEPH mailing list) using RAID0 when passthrough is not supported is BAD (lower performance/IOPS, buggy firmware, some (unknown?) implication on cache, ...).
  • Unknown effect from the cache.
Actions #6

Updated by Timothée Floure almost 4 years ago

  • Status changed from In Progress to Waiting
Actions #7

Updated by Nico Schottelius 3 months ago

  • Status changed from Waiting to Closed
Actions

Also available in: Atom PDF