Performance Monitoring on Pentium 4 Processor

1
Performance Monitoring on Pentium 4 Processor

• Nidhi
• nidhi.nidhi_at_intel.com
• IA 32 Performance Architect

2
Outline

• Pentium 4 Processor Performance Monitoring
  features
• Implementation
• How Intel uses Performance Monitors
• Limitations
• Open issues

3
Feature Overview

• Counters
• 18 40-bit programmable counters
• Events
• 45 events in various parts of the machine
• Counter increment control
• qualification by current privilege level (O/S,
  USER)
• qualification by hardware thread id
• edge detection
• threshold comparison
• interrupt on counter overflow
• Interface ( x86 instructions to set/read
  counters)
• WRMSR (write machine status register)
• RDMSR (read machine status register)
• RDPMC (read performance monitoring counter)
• RDTSC (read time-stamp counter)

4
Features Overview, cont.

• Cascading
• Second counter begins counting when first counter
  overflows
• For instance, to measure cycles elapsed after the
  first counter overflowed.
• Tagging
• Used to get non-speculative event counts
• Tags micro-ops when they incur an event
• Counts tagged micro-ops at retirement
• Three tagging mechanisms front-end, execution,
  and replay

5
Precise Event Based Sampling

• Mechanism
• User allocates a PEBS buffer in memory
• User programs a counter to tag micro-ops and
  count them as they retire
• When the counter overflows, the Pentium 4
  Processor s retirement logic forces a microcode
  assist just before the next tagged micro-op
• Microcode assist copies the program counter and
  GPRs into the PEBS buffer in memory
• Advantages
• Precise taken at instruction which had an event
• Enables creation of data address profiles and
  locate cache lookup patterns and data relocation
  opportunities

6
Implementation Overview
7
How Intel Uses Performance Monitors

• Intel uses Performance Monitoring for
• Performance Analysis
• Compiler optimizations
• System level optimizations
• Performance and functional debug
• Many tools built for analyzing and collecting
  Performance monitoring counters
• Interval Sampler
• Profiler

8
Performance Analysis

• Interval sampler
• Gives the characteristics of the system
• VTune Performance Analyzer
• Event Profiler
• Gives the distribution of events for the system
  over the whole application run
• Available at http//www.intel.com/software/produc
  ts/vtune/
• Interval Sampler points out which events to look
  for, VTune event profiles then help find the
  function, basic block or the IPs that have the
  performance problem.

9
Limitations

• Not all counters can count all events.
• With hyperthreading, the counters may get divided
  among the logical processors.

10
Open Questions

• Centralized Vs. Distributed?
• Distributed is simpler but less flexible
• Add new events?
• New usage models
• Multicore / Multithread scenarios
• Feedback is welcome! ?