Keyun Cheng

Reading Notes: FAST’19 HeART

Title: Cluster storage systems gotta have HeART: improving storage efficiency by exploiting disk-reliability heterogeneity

Conference (FAST’19): Link

Journal (): Link

Summary

This paper introduces HeART, an online tuning tools for redundancy transition by observations of disk reliability status. HeART significantly reduces the # of disks (storage savings) with the same reliability levels (MTTDL) compared with one-scheme-for-all and replications.

Main Contributions

Details

• Goal: reduce storage overhead by replacing one-fits-all schemes with adaptive redundancy rate
  • The reason why AFR comes into play
• Analysis of AFR
  • Bathtub curves of AFR for six disk models are shown
• How to use AFRs?
  • HeART groups disks with similar AFRs and a tailor-made scheme for different groups
    • Heterogeneous AFRs
  • How to measure disk reliability? MTTDL
    • One interesting thing: multiple redundancy schemes can achieve same MTTDL, but long schemes (high k) with high reconstruction costs - there is a tradeoff how to select redundancy schemes?
• Challenges that HeART try to addresses:
  • Online redundancy transition for different groups
    • time required, I/O (not considered)
  • Accurately detects AFRs for different groups
    • How to filter out abnormal AFRs?
• Designs
  • Online change point (bathtub point change significantly) detection
    • Should use prior information about HDDs at start
    • Standard sliding-window based change point detector
    • Periodically trigger point detection
  • Abnormally detection of reliability data stream from HeART (to filter out the abnormal data): use RRCF algorithm (I didn’t look into detail)
• Evaluation
  • Over Blackbaze dataset of 6 disk groups
  • Mostly focus on the functionalities
    • Identify the useful life periods -> MTTDL
    • How HeART suggests the coding schemes with best storage savings?
      • Start from (9,6) -> depend on disk AFRs among groups -> also short codes (like (24, 20) and (21, 17))
  • Sensitivity analysis
    • how the system judges the flat period of disk groups?
    • AFR buffer: the period (or padding called in the paper) added after the infancy of a disk life (in the baththub curve)
      • Larger buffer size: conservative for the low part of AFR

Strength

• Probably the first system work on the disk AFR analysis (I don’t see others for now)
  • which gives strong evidence for the motivation of redundancy transition: disk failure rate highly varies and the baththub curve suggests that we can adapt the redundancy rates with convertible codes
• HeART can suggest how to adapt the redundancy scheme with init code (9,6)
  • It’s based on the analysis disk’s real status (AFR) instead of man-made
• The designs are mostly non-system
  • Based on reliability analysis (MTTDL)

Weakness

• What’s the role of HeART in real systems? This paper suggests that the approach is online, but didn’t show the exact performance of HeART in real systems as a reliability adaptor.