We use the scripts to analyze the predictability of server failures due to DRAM errors as well as the impacting factors on server failures.

Python 3: please install numpy and pandas + pip3 install -r requirement.txt

• The raw data is stored under ../data/
  • python3 measurement.py ../data/
• The results will be stored under ./result/

• overall_distribution: output the percentage of servers with CEs and percentage of servers with server failures per month in Finding 1.
  • Results: overall_distribution.txt
• predictable_analysis: output the relative percentage of predictable server failures for different prediction windows in Finding 2.
  • Results: predictable_analysis.txt
• num_ce_analysis: output the average number of CEs for different types of server failures with different prediction windows in Finding 3.
  • Results: num_ce_analysis.txt
• mtbe_analysis: output the median mean time between errors (MTBE) per predictable server failures for different types of failures with different prediction windows in Finding 4
  • Results: mtbe_analysis.txt
• frac_failure_per_component: output the relative fraction of predictable servers that associated with different memory subsystem component failures for different types of failures when the prediction window is five minutes in Finding 5
  • Results: frac_failure_per_component_5m.txt
• frac_ce_per_component: output the relative fraction of CEs that are associated with different memory subsystem component failures for different typs of failures whehn the prediction window is five minute in Finding 5
  • Results: frac_ce_per_component_5m.txt
• hardware_configuration_impact_analysis: output the relative percentage of predictable server failures breakdown by different hardware configures factors in Findings 6-8
  • Results:
    • DRAM_model_breakdown.txt for DRAM models
    • DIMM_number_breakdown.txt for number of attached DIMMs per server
    • server_manufacturer_breakdown.txt for server manufacturer
• read_scrubbing_analysis: output the average number of read errors and scrubbing errors per predictable server failures for different types of server failures when the prediction window is five minutes in Finding 9
  • Results: read_error_mean.txt and scrub_error_mean.txt)
• hard_soft_analysis: output the average number of hard errors and soft errors per predictable server failures for different typs of server failures when the prediction window is five minutes in Finding 10.
  • Results: hard_erorr_mean.txt and soft_error_mean.txt

Please email to Zhinan Cheng ([email protected]) if you have any questions.