Configurable Cache Subsetting for Fast Cache Tuning

1
Configurable Cache Subsetting for Fast Cache
Tuning

• Pablo Viana1, Ann Gordon-Ross2, Eamonn Keogh2,
• Edna Barros1, Frank Vahid2

1Computer Science Institute CINFederal
University of Pernambuco - UFPE,
Brazil 2Department of Computer Science and
Engineering University of California, Riverside
This work was supported in part by the Capes
Foundation BEX 1366/04-1
2
Outline

• Introduction
• Configurable cache tuning
• Cache configuration space
• Configuration space subsetting problem
• An exhaustive analysis
• A heuristic to subset the configuration space
• Applying the subsetting method to a 2-level cache
• Future work on other configurable platforms

3
Introduction

• Caches can improve system performance. However,
  caches are power hungry.
• Thus, the cache is a good candidate for
  optimization

Power analysis of ARM970TS. Segars (ISSCC, 2001)
4
Motivation

• Tuning cache parameters (Total size, Line size,
  Associativity) to an application can reduce
  energy by 60 on average.

Microprocessor
Cache
Energy
Main Memory
Possible Cache Configurations
5
Cache Configuration Tuning
Application 1
Application 2
Application 3
Application 4
Application 5
...
Application N
Cache Configuration Space
6
Related Work

• Configurable caches
• Soft-core processors
• ARM
• MIPS
• Tensillica
• etc.
• Hard-core processors
• Motorola MCore (Malik, ISLPED00)
• Albonesi (MICRO00)
• Zhang (ISCA03)
• Configurable cache tuning
• Mostly done manually in practice
• Sub-optimal
• Time-consuming

7
Related Work

• Automated Methods for Cache Tuning
• Methods and heuristics for Design Space
  Exploration
• Single-level caches (Tens of configurations)
• Platune (Givargis TCAD02, Palesi CODES02)
• Zhang (RSP03)
• Two-level caches (Hundreds to thousands of
  configurations)
• Tcat (Gordon-Ross, DATE04)

Level 2
Level 1

Total size Line size Associativity
2500 configs
Total size Line size Associativity
Say 50 configs.
Say 50 configs.
8
Problem Context

• Do we really need such a large number of cache
  configurations?
• Could just few carefully-chosen configurations
  adequately cover the large configuration space?
• A smaller configuration space would be easier to
  explore (via simulation-based or dynamic tuning)

9
Problem Context

• Potential scenario
• A configurable microprocessor vendor pre-selects
  a subset of configurations for each particular
  application domain.
• The user selects the most appropriate domain for
  a target application and examine only the
  pre-selected subset.

10
Cache Configuration Tuning
Energy to run a given application on different
configurations
11
Cache Configuration Subsetting
Application 1
Application 2
Application 3
Application 4
Application 5
...
Application N
Cache Configuration Space
Near optimal cache tuning Average energy
increases
12
Cache Configuration Subsetting

• Problem Definition
• Identify the subset of configurations that
  adequately covers the configuration space for the
  application domain.
• We state that p configurations are necessary to
  cover a space of m points.
• Exhaustive Approach
• Select the subset of p configurations from the
  space m
• Criterion for selection
• Choosing the subset which keep the average
  energy of the tuned cache nearest to the optimal.

13
Experimental Setup

• Configurable cache architecture for initial
  experiments
• Our target configurable cache architecture is
  based on Zhang/Vahid/Najjars Highly-Configurable
  Cache Architecturefor Embedded Systems, ISCA
  2003
• We set the base cache to suport the following 18
  configurations
• Energy model for estimation

14
Exhaustive Approach

• Subsets of size p18 down to 1 were chosen
  through the exhaustive evaluation of the average
  energy increase.

15
Exhaustive Approach

• Good results, but exhaustively determining the
  subsets requires too much computation

• For m18 and 1 lt p lt 18 it gives us 262,143
  possible combinations.
• Thats too expensive!
• We need a more efficient way to find the subset
  of configurations...

16
Looking for a Heuristic

• First attempt
• Choose the p configurations that offered optimal
  energy for the largest number of applications.

Energy increase
17
Looking for a Heuristic

• Second attempt
• Hierarchical clustering according to the
  similarity between configurations on their
  average energy savings.

Energy increase
18
Similar Problem

• We found a problem similar to our subsetting
  puzzle.
• The problem of segmenting time series.

19
Similar Problem

• The method proposed by Keoghs can be applied, to
  reduce the number of primary color for displaying
  an image
• (IEEE Int. Conf. On Data Mining, 2001)
• Color/nuances of the image are merged according
  to the associated error/difference in the final
  image.

20
Adapting to the Subsetting Problem

• Similarly, we may iteratively discard
  configurations by merging them.
• By merging two configs cj and ck into ck means
  that all applications which were tuned by cj, now
  use ck.

e(cj,ai)
21
Keoghs Heuristic

• All the possible merges of two adjacent
  (neighboring) configurations in the space are
  evaluated to find the best merging.

c7
c1
c7
c2
c13
c1
22
Keoghs Heuristic

• Accuracy of the results

23
Comparing the Results

• We went on merging the configuration space while
  the energy increase remains under 5 (4 configs).

24
Two-level Configurable Cache
Average energy increase 3.36
25
Where Else this Method Could be Applied?

• Other parameterized IP cores
• Buses, I/O devices (Word size, bandwidth, etc).
• Parameterized Platforms,
• Processors functional units, cache, bus,
  multiplier, IPs.
• Optimization for low energy, area, performance
  and others.

26
Conclusions

• The configurable cache subsetting problem was
  presented
• A Data Mining algorithm for segmenting time
  series was adapted to tackle the cache subsetting
  problem
• Subsetting the Two-level cache configuration
  space, keeping the average energy increase under
  5 takes
• Keoghs heuristic less than 1 minute
• Exhaustively around 53 hours.
• Thanks to the versatility of the proposed method,
  our next step is to apply it for Platform
  customization.

Thank you !