Client Device Affinity in HPSS 7'1

1
Client / Device Affinity in HPSS 7.1

• David Fisher
• Lawrence Livermore National Laboratory
• This work performed under the auspices of the
  U.S. Department of Energy by Lawrence Livermore
  National Laboratory under Contract
  DE-AC52-07NA27344.

2
Client Affinity

• Definition
• Makes local file transfers convenient.
• The HPSS 7.1 client file system affinity feature
  influences the selection of HPSS storage media so
  that Movers that have access to the clients file
  system will be used to read or write the HPSS
  media.

3
Client Affinity

• Definition
• Employed only when the client provides file
  system information in the read or write IOD, and
  when HPSS is configured to support it. It is off
  by default.

4
Client Affinity

• Definition
• The Mover will opens the users file directly and
  transfers to/from the HPSS storage device without
  transferring the file over the network.
• Striped transfers employ multiple Movers, each of
  which opens the users file and transfers
  directly to/from the file and the HPSS stripe
  element.

5
Client Affinity

• Two Kinds of Client Affinity
• Client/tape affinity
• Client/disk affinity

6
Client Affinity

• Client / Tape Affinity
• Not fully designed
• May not be available in HPSS 7.1.0.0
• Tape write
• Core Server selects a tape to write that can be
  mounted on a tape drive operated by a Mover that
  has access to the clients file system.
• Tape read
• Core Server mounts the tape on a tape drive
  operated by a Mover that has access to the
  clients file system.

7
Client Affinity

• Client / Disk Affinity
• Completely designed, mostly implemented, and
  partially tested.
• Disk write
• Core Server restricts its selection of disk
  volumes to create the user file on volumes that
  are operated by Movers that have access to the
  users file system.
• Disk read
• Not implemented for disk read because the HPSS
  disk file might have to be moved from one volume
  to another to land on a suitable volume.
• But may become part of file staging.

8
Client Affinity

• Client / Disk Affinity
• Repack
• When disks are repacked, files are moved to
  volumes that have the same client file system
  affinity as the source volumes.

9
Client Affinity

• Describing user file systems to HPSS
• HPSS 7.1 contains a new metadata table, the
  FILESYSTEMS table.
• The FILESYSTEMS table maps host names and file
  system mount points to arbitrarily selected File
  System Ids.
• Allows file systems to be mounted at different
  locations in client and mover systems.
• We will provide a utility program to set up and
  maintain this metadata table.

10
Client Affinity

• Describing user file systems to HPSS
• Example of the table
• FSID Host Name Mount Point
• 1 Host2 /root/mp1/
• 2 Host1 /root/mp2/
• 2 Host2 /root/mp2/
• 3 Host1 /root/mp3/

11
Client Affinity

• How does it work?
• Tape Affinity
• Hmmmmnot sure yet. Well probably need some
  additional metadata to map tape drives to movers
  and file systems.
• Beyond that, things are foggy right now.

12
Client Affinity

• How does it work?
• Disk
• The Core Server maps disk volumes to disk movers
  and maps file systems, described in the
  FILESYSTEMS table to movers.
• Each mover in the Core Servers catalog has a
  list of file system IDs associated with it, and a
  list of disk volumes it operates.
• When the file is created on disk, only the disks
  operated by the appropriate movers are considered.

13
Device Affinity

• Definition
• The device affinity feature is defined for disks
  only.
• Its purpose is to evenly spread files across the
  disk volumes and to take I/O workload into
  consideration when selecting a disk volume for a
  file.
• Always on, cannot be disabled.
• Requires no initial configuration.

14
Device Affinity

• Design Selecting a disk for a file
• Idle Movers are chosen before busy Movers.
• Then, least recently allocated Movers are chosen
  first.
• Idle disk volumes are chosen before busy volumes.
• Then, least recently allocated volumes are chosen
  first.
• Both read and write operations are monitored.
• Our first implementation uses a very simple
  scheme for accounting for disk read and write
  operations. We expect this aspect of the design
  to evolve as we learn how it performs in the real
  world.

15
Device Affinity

• Expected benefit
• Even distribution of disk files across movers and
  volumes.
• Even distribution of disk I/O workload across
  movers and volumes.
• Movers and devices should not be accidentally
  choked with work.
• Disk bandwidth utilization should ramp up more
  smoothly.

16
Device Affinity

• Actual benefit
• Well let you know.

17
Device Affinity

• Transitioning to this feature
• Designed to be automatic. No configuration
  required.
• Some sites may need to adjust the layout of
  striped disk volumes on Movers.
• Documentation will cover this in detail.
• Any changes that should be made can be gradual.

18
Go Boilers!