The HP AutoRAID Hierarchical Storage System

1
The HP AutoRAID Hierarchical Storage System

• John Wilkes, Richard Golding, Carl Staelin, and
  Tim Sullivan
• Hewlett-Packard Laboratories
• Presented by Sri Ramkrishna

2
HP AutoRAID Hierarchical Storage System Overview

• Two level storage hierarchical implementation
  inside a single array controller.
• Consists of a mirror copy for fast storage
• RAID5 storage for slower storage
• Seamlessly moves data from one to the other.

3
Overview

• What is a RAID system?
• RAID is Redundant Array of Disks
• Usually comes in RAID3 or RAID5
• RAID3 is some number of disks with one disk
  dedicated for parity
• RAID5 is some number of disks, where each 1 block
  on each disk creates a stripe, plus a parity
  block.
• Requires an array controller
• Mirror (possibly RAID1)
• Two copies on two disks. Generally faster.

4
Disk Arrays

• Problem is disk arrays are hard to use.
• Requires understanding of the disk load
• If you mess up, its expensive to fix.
• System performance becomes degraded.
• Have to move data off to another storage.
• Adding new capacity, or new disks require you
  move data off and then restore.

5
Hierarchical Storage the Solution

• Combine the performance of mirrored disks with
  cost-capacity benefits of RAID5.
• Constraints
• Active data must change slowly
• Can be implemented three ways
• Manually
• Error prone
• In the filesystem not particularly portable
• In smart array controller
• How HPs solution is done.

6
Important Features

• Mapping to allow transparent migration of disk
  blocks.
• Mirroring and RAID5
• Adaption to Changes in Amount of Data Stored.
• Starts empty, data stored in mirrored space till
  full then gets migrated to RAID5.
• Has a fine granularity of 64k unit when moving
  data between mirrored and raid5.
• Hot pluggable disks, fan, power supplies, and
  controllers.

7
Features continued

• On-Line Storage capacity expansion
• Can add up to 12 disks transparently
• New disks are easily added
• Active hot spare
• Simple administration
• Log-structured RAID 5 writes.

8
AutoRAID Details

• Similar to regular RAID array
• Set of disks, intelligent controller, caches for
  staging data
• Physical layout consists of
• Physical Extents (PEXes)
• 1M in size
• Consists of 128K segments
• Segments are either part of mirrored set, or
  RAID5
• Physical Extent Group (PEG)
• Stripe of PEXes.
• PEXs allocated such that they distribute the
  load across all disks.
• At least on three disks
• Are assigned to either mirrored or raid5 or
  unassigned

9
(No Transcript)
10
(No Transcript)
11
Logical View

• To machines, AutoRAID presents storage as logical
  64K pieces called Relocation Blocks (RBs)
• When a new LUN (Logical Unit Number) is created
  or is increased, its address space is mapped to
  unto a set of RBs
• LUNs are the logical address for each individual
  drive in a disk array.
• Allocation occurs on write.
• Each PEG can hold a number of RBs
• Is a function of the size of the PEG

12
How it works

• Host initiates a read or write operation to the
  disk array.
• Reads can be cached by the array which can be
  pretty fast.
• Writes are more complicated

13
AutoRAID Writes

• Has an non-volatile NVRAM
• Host can load request into the NVRAM, once
  complete, host believes its request is done.
• Some policies might wait for for additional
  writes to batch the writes together
• NVRAM is flushed, and a background write is
  initiated.
• If the data exists in mirrored space, the data is
  written there.
• Otherwise, the data is promoted to mirrored space
  since its now active and then written.

14
Promotions

• Migration code is called to move data from RAID5
  space to mirrored space
• If no space is left in mirrored space, some space
  is demoted down to RAID5 space.
• There are some tricky situations where there
  might be a catch 22 situation that needs to be
  handled.

15
Reads and Writes

• Reads and Writes in Mirrored space is simple.
  Reads pick one of the copies and reads it.
  Writes are done by writing to both disks. Write
  is complete when both disks are written to.
• Reads in RAID5 space is pretty straightforward.
• Writes in RAID5 is more complicated.

16
RAID5 Writes

• RAID5 storage is layed out like a log
• Means that RBs that move from mirrored space is
  appended to RAID5 storage PEG.
• Depending on whether it has free slots of course.
• RB writes can be done in two ways
• Per RB
• Generates two disk writes, one for data one for
  parity
• Batched writes
• Waits for all the RBs in a stripe is written.
• Only has one parity write
• Commonly used in most RAID5 implementation

17
CompactionsHoles, Garbage Collection

• Demoting and promoting causes holes in mirrored
  space
• Added to free list
• Can be reused for promotions from RAID5 space
• Can also be used to fill holes to free up a PEG,
  so it can be used in RAID5 storage.
• Same problem in RAID5 space
• Called garbage collecting
• Holes cannot be filled, but must be cleaned up.

18
Migrations/Balancing

• Migrations to RAID5 space from Mirrored Space
• Rbs are selected by Least Recently Written (LRU)
  selection.
• Done in the background
• Balancing
• When new drives are added, migration is done to
  balance the performance.

19
Testing Setup

• Baseline configuration was 12 disk system with
  one controller and 24MB of controller data cache.
• Connected to HP 9000/ K400 system with one
  processor and 12 MB
• Compared against Data General CLARiion disk array

20
Performance Results

• AutoRAID vs RAID Array vs JBOD-LVM
• OLTP show that AutoRAID out performs RAID, and
  3/4th of JBOD-LVM
• JBOD-LVM
• JBOD means Just a Bunch of Disks
• Writes were slower than JBOD-LVM because mirrored
  writes were slower than JBOD-LVM.

21
(No Transcript)
22
(No Transcript)
23
Some Notes

• Increasing the speed of the disks, improves the
  backend peformance.
• Improving transfer rate is more important than
  rotational latency

24
Summary

• HP AutoRAID is very easy to use
• Sysadmins are able to add disks, and do various
  tasks without having to worry about whether the
  disk layout is correct.