HOWTO Systems Research

Systems Research

• Systems research: solving problems in computer systems
  • Performance, energy consumption, cost, security, etc.
• Worst research: try to solve non-existing problems
  • Make the performance of a keyboard device driver 10x faster
  • Does the performance of the keyboard device driver really matter?
• Finding real, important problems are critical

How to find real, important problem

• Read recent conference/workshop papers
  • What problem area is important
  • What problems are already solved and what are not solved yet
  • What are commonly used techniques in this field
  • What are commonly agreed ways to evaluate a proposed methodology
• However, a paper is the result of at least one-year effort
  • You are already two years behind than the authors

How to find real, important problem as early as possible

• Attend top-tier conference and chat with researchers
  • What are you working on these days?
  • Good to get a sense of research direction
  • If you are lucky, you can find a new, interesting problem
• Build a prototype of an existing system
  • Best way to have deepest understanding of an existing system
  • It takes time so it is risky if there is no new, interesting problem
• Benchmarking
  • Economic way to find new, important problems

How to use benchmarking to discover new problems

• Run existing benchmark in new setting
  • E.g., run mosbench on 200 core machine -> Can we find any new bottleneck?
  • E.g., run mosbench on docker, virtual machine, or unikernel -> Does new isolation mechanism introduce new overhead?
  • E.g., run FxMark on an extremely-fast NVM or a slow SMR drive -> How does performance characteristics of storage device affect performance and scalability of storage stack?

How to use benchmarking to discover new problems

• Compare existing approaches with the same benchmark
  • Can have deeper understanding of existing approaches beyond what described in the paper
  • Can draw design principles of a new approach

What are important in benchmarking

• This is not an one time job
  • Automation, automation, automation!
• Without automation, this is not a reproducible science.
  • We human do make mistakes.
  • Automation is one of way to avoid mistakes (or make the same mistakes consistently)
• If we are copying log files here and there, this is the sign that we should automate

What are important in benchmarking

• Automate as much as possible
  • Run all necessary combinations of experiments
  • Generate graphs
  • (Email generated graphs automatically)
• It does affect our productivity a lot!
  • Run a benchmark script before having meeting (or going to bed, leaving office, etc)
  • After meeting, check graph and devise what to run next or what to profile
• bash, awk, and python are our friends

Examples of good benchmarks

• Mosbench (or our forked vbench)
  • https://pdos.csail.mit.edu/archive/mosbench/
  • https://github.com/sslab-gatech/vbench
• FxMark
  • https://github.com/sslab-gatech/fxmark

Tips for more meaningful benchmarking

• Make your results easy to compare
  • ## transactions/msec instead of ## transactions
  • Abort ratio instead of ## aborts
• Double check benchmark parameters
  • number of initial elements in a hash table
  • update ratio
  • random distribution: uniform vs. zipf

Tips for more meaningful benchmarking

• Know your hardware
  • Take care of CPU topology in pinning threads
  • Check /proc/cpuinfo
  • Or use cpu-topology.py, cpu-sequences, or set-cpus
• Avoid well-known bottlenecks
  • Use scalable memory allocator, such as jemalloc or tcmalloc
• Use git, tmux, and gnuplot, seriously