Architectural Support for SemanticallySmart Disk Systems

1
(Architectural Support for)Semantically-Smart
Disk Systems

• April 29, 2003
• Jarrod Lewis
• University of Wisconsin-Madison

2
Objectives

• Semantically-Smart Disk Systems (SDS)
• Disk processing for SDS
• Architecture brainstorm

3
Overview Motivation

• Observe storage systems important
• Want better storage!
• Problem fixed interfaces
• Hard to change OS, file systems
• Solution innovate around interfaces
• Infer information

4
Overview Approach

• Where to innovate?
• At lowest level (disks)
• The semantic gap
• Infer higher level semantic information
• Exploit in lower levels
• Semantically-Smart Disk Systems (SDS)
• Reverse engineer disk traffic
• E.g, Distinguish meta-data from data
• Handle types differently

5
Overview Applications

• Improve performance
• Smarter layout
• Smarter file caching
• Improve availability
• D-GRAID
• RAID that degrades gracefully
• Structural caching
• Explicit journaling of metadata
• Improve security
• Secure delete

6
Overview Issues

• Traditional RAID processing
• Parity calculation
• Simple prefetching
• SDS an evolving workload
• Block differencing
• Bitwise diff between 4kB chunks of data
• Hash tables
• E.g, to distinguish type of (overloaded) data
  blocks
• Deferred work
• Save block movement for idle time

7
Outline

• Overview
• Motivation
• Approach
• Semantically-Smart Disk Systems
• Architectural Issues

8
Motivation

• Storage critical to useful computer systems
• Want improvements!
• Performance, availability, security, etc
• But
• OS and storage interfaces lack explicit hooks
• So
• Improve storage with implicit hooks

9
Outline

• Overview
• Motivation
• Approach
• Semantically-Smart Disk Systems
• Architectural Issues

10
Approach

• Reverse engineer disk traffic
• Determine semantics
• Behind narrow interface (SCSI)
• Build new functionality
• Block-based interfaces NOW and LATER
• SCSI interface lives on
• iSCSI (SCSI over IP) emerging
• Processing support?
• Data movement vs. computation
• Block differencing
• Hash tables

11
Outline

• Overview
• Motivation
• Approach
• Semantically-Smart Disk Systems
• Architectural Issues

12
Outline

• Overview
• Motivation
• Approach
• Semantically-Smart Disk Systems
• Architectural Issues

13
SDS Hashing Space Overheads
14
Block differencing

• Completely independent computations
• SIMD? Vectors?
• Distribute slices of blocks to multiple
  processors?
• Lots of block storage needed
• How to design memory hierarchy?

15
Hardware hash tables

• Terrible locality
• Bad for traditional memory hierarchies
• Lots of storage needed
• What memory structures to use?
• Distribute hash table?
• Parallelize lookups
• Think of disks as multiple cache banks?

16
Deferred work

• Queue up work (block movements)
• Wait for idle time
• Buffering overhead?

17
Other ideas/issues?
18
Thank you for coming.