| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102 |
- Introduction
- What is it?
- A bunch of speed of light benchmarks,
- not MP, not throughput, not saturation, not stress tests.
- A microbenchmark suite
- Measures system performance
- Latency and bandwidth measurements
- Measurements focus on OS and hardware
- What is delivered to the application
- Not marketing numbers
- Benchmark performance predicts application performance
- Results for which systems?
- Sun, SGI, DEC, IBM, HP, PCs
- Useful information to whom?
- Performance engineers, system programmers, system architects.
- Motivation
- What are we measuring?
- Control / latecy operatins
- Bandwidth operations
- What aren't we measuring?
- Basic MIPS & MFLOPS. XXX - not unless I do it right.
- What can I learn?
- Cost of operations
- ****Operations per time unit****
- Compare speed of alternative paths (e.g. mmap vs. read)
- Performance problems = f(bw issues + latency issues)
- Give at least two examples
- NFS control & data: UDP lat, proc lat, & various BW metrics
- Oracle lock manager: TCP lat
- Verilog: mem lat
- AIM: fs ops XXX -ask Scott about pipes.
- Knowing the speeds of primitives can provide speeds of apps.
- An example here would be nice.
- Outline
- Describe benchmark
- Give results from current machines
- Discuss results
- Future changes, enhancements, etc.
- Tutorial on benchmarks
- For each metric
- what is it?
- why is it being measured?
- How is it measured?
- Measuring subtlities
- Interpreting the results
- Latency
- Process stuff
- networking stuff
- file system stuff
- memory stuff
- whatever
- Bandwidth
- networking
- file system
- memory
- Results
- Tabular results - XXX update that table to reflect the newer metrics
- Graphs of memory latency & context switches
- Discussion
- Memory stuff
- Maybe contrast AIX with the $100K IBM
- uniprocessor w/ killer memory perf and point out
- that it is the memory that is making AIX go
- fast, it certainly isn't AIX. A more politic
- observation would be that systems with good
- memory performace tend to have good system
- performance; the point being to shift people's
- attention to system performance, especially
- memory subsystem, as opposed to processor mips.
- Comparisons
- Maybe look at the table and draw attention to
- really good and really bad numbers for various
- platforms (like Linux' context switch time,
- Linux fs ops, solaris syscall, process stuff,
- 990 memory BW).
- Graphs
- A graph showing a range of really fast to really slow ops, all on the
- same graph. Do bandwidth stuff normalized on MB/sec.
- Carl sez: show both ops/sec and cost/op on two graphs.
- A graph showing processor slow down due to memory misses, assuming
- each instruction misses. Maybe a graph that shows # of clocks
- (or better yet, # of instructions - think super scalar) that you would
- have to have between each memory miss in order to run at the clock
- speed.
- War stories
- Sun page coloring bug
- SGI page coloring bug
- SGI hippi bug - XXX ask Thomas
- Sun bcopy bug
- Lmbench [optional?]
- how to get lmbench
- how to compile
- how to run
- how to show results
- Future work
- More hardware stuff - better latency measurements (write lat,
- cache to cache latency).
- add throughput & saturation measurements
- TODO
- get some similar papers for comparison
- Someday I need reasonable I/O benchmarks to show off good
- big SMP machines like Challenge.
|
