Pierfrancesco Foglia

1
Classification and Evaluation of Sharing Overhead
of Commercial Applications on Shared-Memory
Multiprocessors

• Pierfrancesco Foglia
• foglia_at_iet.unipi.it

2
Why Commercial Applications?

• There is a large amounts of studies targeting
  scientific applications (based on the SPLASH
  benchmark suite) but few concerning commercial
  applications
• Commercial applications (DBMS, WEB and
  Multimedia) are becoming the most utilized
  workloads for multiprocessor servers and have
  quickly surpassed the scientific/engineering
  applications
• DBMS two classes of workload
• OLTP ( TPC-B benchmark with PostgreSQL DBMS)
• DSS ( TPC-D benchmark with PostgreSQL DBMS)
• WEB applications
• WEB Server with search engine ( a sequence of GET
  on an Apache server with TPC-B query simulating
  the search engine)
• Multimedia Station
• A collection of audio and video applications,
  java (multithreaded) applets etc.

3
Our Workloads (I)

• performance is effected by the number of
  processes running on the server
• we want to characterize memory performance
  during different phases of execution
• of the benchmark ( it is a dynamic situation)

4
Our Workloads (II)
5
Our Workloads (III)
6
Our Workloads (IV)
7
Our Workloads (V)

• TPC-B workloads present the same statistics by
  varying the number of processes (same queries
  executed in the benchmark)
• TPC-D workloads present the higher number of
  writes and shared writes (accesses to larger data
  structures)
• SWEB workloads present the higher number of
  distinct block, code accesses and almost no
  shared writes (server and client only communicate
  the server state)

8
Our Aims

• Evaluation of sharing overhead of commercial
  application in shared memory multiprocessors
• Evaluation of passive sharing and its overhead in
  these systems (it is introduced by the scheduler,
  necessary to achieve load balancing)
• Evaluation of a coherence protocol (PSCR)
  designed to eliminate passive sharing
• All the previous points applied to multithreaded
  system

9
Our Methodology

• Trace-driven simulation by the mean of Trace
  Factory environment
• Our on line version of Hydes algorithm for the
  evaluation of degenerate sharing ( Hyde, Fleisch,
  Degenerate Sharing, 1994 International Conference
  on Parallel Processing)
• Extension of the algorithm to the passive sharing

10
Hydes Algorithm (I)

• It assumes infinite cache size
• It is off line (after simulation)
• True Sharing it occurs when one site reads a
  value written by another site. In this case
  there is a real communication between the two
  sites and this implies that the coherence
  transmission is necessary
• pattern we had a coherence transaction and the
  following pattern of accesses
• False Sharing

11
Hydes Algorithm (II)

• We need also to specify the patterns which
  determine a coherence transaction (off line
  algorithm)
• Overwrite Sharing two writes from different
  sites without any read
• Degenerate Sharing all the sharing that is not
  True Sharing

12
Our Extension

• During the simulation we can detect when a
  coherence transaction happens
• We can detect replacement activity (for finite
  size cache)
• Passive Sharing a coherence transaction on a
  private block (all the other sharing concerns
  shared blocks)

13
Some Results
14
Some Results
15
Some results
16
Some Results