Cache-Conscious Structure Definition

1
Cache-ConsciousStructure Definition

• By Trishul M. Chilimbi, Bob Davidson, and James
  R. Larus
• Presented by Shelley Chen
• March 10, 2003

2
Motivation

• Processor-Memory Performance Gap
• Growing at 53 per year!
• Chilimbi and Larus previous work
• Placed objects with high temporal locality in the
  same cache block
• Works best with objects lt ½ cache block
• Current paper proposes techniques for larger data
  structures

3
Improving Cache Performance
4
Two Cache-Conscious Definition Techniques

• Structure Splitting
• Split large data structures into two smaller
  structures
• Field Reordering
• Group fields in a structure with high temporal
  locality into the same cache block

5
Structure Splitting

• Split large structures into two smaller
  structures
• hot structure contains frequently accessed
  fields
• cold structure contains rarely accessed fields
• Allow more hot structures to fit into the cache
• Has been done manually in the past, this paper is
  first to automate

6
Class Splitting Overview
7
Experimental Setup

• Ran compiled programs on Sun Ultraserver E5000
• 2 GB of memory
• L1 dcache, 16 KB DM, 16 byte blocks
• L2 cache, 1 MB DM, 64 byte blocks

8
Results Structure Splitting
Reduces L2 miss rates by 10-27 improves
execution time by 10-20
9
Field Reordering

• Logical ordering of the program is not usually
  consistent with its data access patterns
• Frequently accessed fields may be coded next to
  rarely accessed fields, putting them in the same
  cache block
• cause excessive cache misses
• Reorder field definitions of structure
• fields with high temporal affinity in same cache
  block

10
bbcache

• Tool that produces structure field reordering
  recommendations
• Construct a database containing both static and
  dynamic information about the structure field
  accesses
• Process database to construct field affinity
  graphs for each structure
• Produce the structure field order recommendations
  for the affinity graphs
• Attempts to group fields with high temporal
  affinity into the same cache block

11
Experimental Setup

• 4 processor 400MHz Pentium II Xeon system
• 1MB L2 cache/processor
• 4GB memory
• Ran TPC-C on Microsoft SQL Server 7.0 to collect
  traces as input to bbcache
• Chose 5 structures which showed largest potential
  from the benefits of reordering

12
Results Field Reordering
Modified SQL Server was consistently better by
2-3

• Cache block pressure S (b1, ,bn)/n
• Cache block utilization S(f11, , fnbn)/ S(b1,
  ,bn)

13
Conclusion

• 2 techniques to improve cache performance
• change the internal organization of fields in a
  data structure
• Structure splitting
• Field reordering
• Structure Layouts better left to compiler
• Easier to determine field access order

14
Questions?
15
Class Splitting Algorithm
16
Structure Access Database
17
Building Field Affinity Graphs