Memory System Characterization of Commercial Workloads

1
Memory System Characterization of Commercial
Workloads

• Luiz Andre Barroso, Kourosh Gharachorloo, and
  Edouard Bugnion
• Presented by Jerry Wu

2
Introduction

• Motivation
• Commercial workloads has become the largest
  market segment for multiprocessor servers
• Design of these systems has yet to keep up with
  the pace of changes in the market
• Lack of commercial workload performance
  requirements
• This paper presents performance studies of three
  classes of commercial workloads

3
Complications

• Lack of availability and restrictions
• Lack of easy access to commercial database
  engines
• Large scale
• Large hardware cost
• Complexity
• Non-trivial OS-I/O interactions, lack of source
  code
• Moving target
• Commercial database engines improve at a very
  fast pace

4
Commercial Workloads

• OLTP Workload
• Modeled after the TPC-B benchmark
• Models a banking system
• DSS Workload
• Modeled after the TPC-D benchmark
• Simulates the decision support system for a
  supplier
• OLTP and DSS workloads run on Oracle Database
  Engine
• Web Index Search Workload
• State-of-the-art search engine AltaVista (No
  Google yet)

5
Methodology

• Monitoring
• OLTP and DSS benchmarks were run on Alpha 21164
  using Oracle
• Utilized IPROBE monitoring tool to access event
  counters
• Simulation
• Used an Alpha port of SimOS
• Key issues
• Amount of physical memory required
• Bandwidth requirement for the I/O
• Total runtime

6
Monitoring Results

• OLTP
• Importance of hits and misses to secondary caches
  and latency of dirty misses
• DSS and AltaVista
• Hits in secondary on-chip cache is the only
  significant memory component

7
Simulation Results

• Uses an Alpha port of SimOS
• Observations on OLTP
• Small kernel component
• Benefit from larger cache and higher
  associativity
• Cache and memory system stalls have large effects
• Idle time increases with bigger caches
• Higher processing rates results in less demand on
  I/O
• Small fraction of communication in Oracle due to
  false sharing

8
Cache Hierarchy Performance

• Primary cache miss important for OLTP, but more
  so for DSS
• OLTP and DSS have very different cache
  performance
• Large on-chip cache captures most of the misses
  in DSS, but not in OLTP
• False sharing increases for increasing cache line
  size
• Replacement and instruction miss rate not visibly
  effected

9
Questions