Improving Application Performance through Swap Compression

1
Improving Application Performance through Swap
Compression

• R. Cervera, T. Cortes, Y. Becerra and S. Lucas

2
Motivation

• Problem
• Large applications
• Highly-loaded systems
• Laptops
• Laptops usually have less resources than desktops
• Homework
• Home machines are usually smaller than office
  machines
• Solution
• Swapping mechanism
• But IT IS VERY SLOW !!!

3
Objectives

• Increase swapping performance
• Without adding more memory
• Increase swapping space
• Without adding more disk
• Performance is more important than space
• Modify the kernel as little as possible
• Mechanism feasible in user-level libraries

4
Traditional Swapping Path
Process B
Process C
Process A
Process D
Operating System
Swap Device
5
Proposal

• Cache for swapped pages
• If the system uses memory for the cacheThen the
  applications lose this memory
• Problem
• Less fast memory
• More slow memory
• Compression of swapped pages
• More than one compressed page per buffer

6
Compressed Swapping v1.0
Operating System
Cache
Swap Device
7
Hit Ratio and Its Kinds

• There are two kind of cache hits
• Hits due write
• Page recently swapped out
• Hits due read
• Page brought to the cache with another request
• Many more hits due write than hits due read
• Different swap-in and swap-out order
• No spatial locality
• Reads produce interferences in the cache

8
Optimizations

• Different read and write paths
• Reads do not place information in the cache
• Reads use the cache (hits due write)
• Reads do not modify the cache
• Writes place information in the cache
• Writes modify the cache
• Batched writes
• Write many cache buffers to disk at the same time
• Write them contiguously
• A page is never split among two buffers
• Reads perform, at most, one disk access

9
Compressed Swapping v2.0
Operating System
Cache
Swap Device
10
Evaluated Environment

• Environment
• Pentium II - 350Mhz
• 64 Mbytes of memory
• Ultra-SCSI disk, 128 Mbytes swap partition
• Bencmarks

11
Performance Evaluation
6,5

• Parameters
• Cache size 1Mbyte
• Cleaning threshold 50
• Better performance
• Speedups better than 1.2
• FFT
• Speedup 0.96
• Bad compression ratio
• Working set
• Simulator x5
• Good compression ratio
• All pages fit in the cache

12
Cache-size Influence

• No perfect cache size
• Large caches are bad
• Use too much application memory
• Recomended size
• Around 1Mbyte

13
Batched-write Influence

• Small values are bad
• Small writes
• Latency is not hidden
• Large values
• Mono-process bench
• Good
• Multi process bench
• Reads confilct with cleans
• Recommended value
• Around 10

14
New-swap Capacity

• Some improvement achieved
• Allows larger applications
• Depends on
• Compression ratio
• Fragmentation

15
Related Work

• Compressed cache (Douglis93)
• No difference between reads and writes
• Limited batched writes
• Performance gains only for applications with high
  compression ratio
• Single process benchmarks
• MagnaRAM
• Memory compression for Windows
• Not very good results

16
Conclusions

• Simple mechanism to
• Increase performance of large applications
• Increase swap space
• Easy to implement
• Modified
• 6 routines and 2 files
• Added
• 9 routines and 1 include file (.h)
• Easy to port from one version to another
• 15 minutes
• Can be used in out-of-core applications