Architecture - Wide Striping

From Lustre.org

Note: The content on this page reflects the state of design of a Lustre feature at a particular point in time and may contain outdated information.

There are several use cases where Lustre wants to write exceptionally many stripes in files:

1. Major HPC installations may have many hundreds or thousands of OSTs and we need to be able to stripe files over all of them
2. Server Network Striping (SNS) will use parity declustering, resulting in an very large number of objects building up the striped file.

Therefore, wide striping will be a commonly encountered case. The goal is to encode the striping information in a very compact way.

Definitions (see fid-hld)

A pool

defines an un-ordered sets of OSTs and will be used to describe the striping in a manageable way.

fid seq number

part of fully specified FID, contains sequence in which object was created

fid number

part of fully specified FID, contains object id within its sequence

object version

part of fully specified FID, contains object version number

FID

fully specified object identification structure: FID = {f-sequence, f-number, f-version}

FLDB

FID Location DataBase, provides fid sequence to server (OST, MDS) mapping

APIs required

1. Get a consecutive set of fid sequence numbers from the FLDB
2. define an on-disk EA that contains a pool name and other RAID striping parameters, for use as a default directory EA
3. define an on-disk EA that contains a RAID type, raid parameters, a starting fid sequence number, a count of objects over which the object may be striped, a sequence skip count, a single fid number used by this file in all specified sequences, the object version, possibly the pool from which this object was allocated (for future reference)
4. offsets within the file are {lov_offset, stripe_index} = fn(file_offset, raid_type, raid_parameters}
5. individual objects OBJ{0, ..., num_obj - 1} in the file can be located:
   • OST(stripe_idx) = FLDB(seq_start + stripe_idx*seq_skip)
   • OBJ(stripe_idx) = FID{seq_start + stripe_idx*seq_skip,fid_number,obj_version}

Back to Top