1. THE HP AUTORAID HIERARCHICAL STORAGE SYSTEM
J. Wilkes, R. Golding, C. Staelin T. Sullivan
HP Laboratories, Palo Alto, CA

2. INTRODUCTION
must protect data against disk failures: too frequent and too hard to repair
possible solutions:
for small numbers of disks: mirroring
for larger number of disks: RAID

3. RAID Typical RAID Organizations
Level 3: bit or byte level interleaved with dedicated parity disk
Level 5: block interleaved with parity blocks stored on all disks

4. LIMITATIONS OF RAID (I)
Each RAID level performs well for a narrow range of workloads
Too many parameters to configure: data- and parity-layout, stripe depth, stripe width, cache sizes, write-back policies, ...

5. LIMITATIONS OF RAID (II)
Changing from one layout to another or adding capacity requires downloading and reloading the data
Spare disks remain unused until a failure occurs

6. A BETTER SOLUTION A managed storage hierarchy: mirror active data
store in a RAID 5 less active data
This requires locality of reference:
active subset must be rather stable: found to be true in several studies

7. IMPLEMENTATION LEVEL Storage hierarchy could be implemented
Manually: can use the most knowledge but cannot adapt quickly
In the file system: offers best balance of knowledge and implementation freedom but specific to a particular file system
Through a smart array controller: easiest to deploy (HP AutoRAID)

8. MAJOR FEATURES (I)
Mapping of host block addresses to physical disk locations
Mirroring of write-active data
Adaptation to changes in the amount of data stored:
Starts RAID 5 when array becomes full
Adaptation to workload changes:
Hot-pluggable disks, fans, power supplies and controllers

9. MAJOR FEATURES (II)
On-line storage capacity expansion: system switches then to mirroring
Can mix or match disk capacities
Controlled fail-over: can have dual controllers (primary/standby)
Active hot spares: used for more mirroring
Simple administration and setup: appears to host as one or more logical units
Log-structured RAID 5 writes

10. RELATED WORK (I) Storage Technology Corporation Iceberg:
also uses redirection but based on RAID 6
handles variable size records
emphasis on very high reliability

11. RELATED WORK (II) Floating parity scheme from IBM Almaden:
Relocated parity blocks and uses distributed sparing
Work on log-structured file systems at U.C. Berkeley and cleaning policies

12. RELATED WORK (III) Whole literature on hierarchical storage systems
Schemes compressing inactive data
Use of non-volatile memory (NVRAM) for optimizing writes
Allows reliable delayed writes

13. OVERVIEW Processor, RAM and Control Logic Parity Logic 2x10MB/s bus
DRAM Read Cache
Matching RAM
NVRAM Write Cache
Control
Other RAM
SCSI Controller
20 MB/s
Host Computer

14. PHYSICAL DATA LAYOUT
Data space on disks is broken up into large Physical EXTents (PEXes):
Typical size is 1 MB
PEXes can be combined to form Physical Extent Groups (PEGs) containing at least three PEXes on three different disks
PEGs can be assigned to the mirrored storage class or to the RAID 5 storage class
Segments are the units on contiguous space on a disk (128 KB in prototype)

15. LOGICAL DATA LAYOUT
Logical allocation and migration unit is the Relocation Block (RB)
Size in prototype was 64 KB:
Smaller RB’s require more mapping information but larger RB’s increase migration costs after small updates
Each PEG holds a fixed number of RB’s

16. MAPPING STRUCTURES
Map addresses from virtual volumes to PEGs, PEXes and physical disk addresses
Optimized for finding fast the physical address of a RB given its logical address :
Each logical unit has a virtual device table listing all RB’s in the logical unit and pointing to their PEG
Each PEG has a PEG Table listing all RB’s in the PEG and the PEXes used to store them

17. NORMAL OPERATIONS (I)
Requests are sent to the controller in SCSI Command Descriptor Blocks (CDB):
Up to 32 CB’s can be simultaneously active and 2048 other ones queued
Long requests are broken into 64 KB segments

18. NORMAL OPERATIONS (II)
Read requests:
Test first to see if data are not already in read cache or in non-volatile write cache
Otherwise allocate space in cache and issue one or more requests to back-end storage classes
Write requests return as soon as data are modified in non-volatile write cache:
Cache has a delayed write policy

19. NORMAL OPERATIONS (III)
Flushing data from cache can involve;
A back-end write to a mirrored storage class
Promotion from RAID 5 to mirrored storage before the write
Mirrored reads and writes are straightforward

20. NORMAL OPERATIONS (IV)
RAID 5 reads are straightforward
RAID 5 writes can be done:
On a per-RB base: requires two reads and two writes
In batched writes: more complex but cheaper

21. BACKGROUND OPERATIONS
Triggered when array has been idle for some time
Include
Compaction of empty RB slots,
Migration between storage classes (using an approximate LRU algorithm) and
Load balancing between disks

22. MONITORING System also includes: An I/O logging tool and
A management tool for analyzing the array performance

23. PERFORMANCE RESULTS (I)
HP AutoRAID configuration with:
16 MB of controller data cache
Twelve 2.0GB Seagate Barracuda disks (7200rpm)
Compared with:
Data General RAID array with 64 MB front-end cache
Eleven individual disk drives implementing disk striping but without any redundancy

24. PERFORMANCE RESULTS (II)
Results of OLTP database workload:
AutoRAID was better than RAID array and comparable to set of non-redundant drives
But whole database was stored in mirrored storage!
Micro benchmarks:
AutoRAID is always better than RAID array but has smaller I/O rates than set of drives

25. SIMULATION RESULTS (I)
Increasing the disk speed improves the throughput:
Especially if density remains constant
Transfer rates matter more than rotational latency
64KB seems to be a good size for the Relocation Blocks:
Around the size of a disk track

26. SIMULATION RESULTS (II)
Best heuristics for selecting the mirrored copy to be read is shortest queue
Allowing write cache overwrites has a HUGE impact on performance
RB’s demoted to RAID should use existing holes when the system is not too loaded

27. SUMMARY (I) System is very easy to set up:
Dynamic adaptation is a big win but it will not work for all workloads
Software is what makes AutoRAID, not the hardware
Being auto adaptive makes AutoRAID hard to benchmark

28. SUMMARY (II) Future work includes: System tuning especially
Idle period detection
Front-end cache management algorithms
Developing better techniques for synthesizing traces