Computer Science Thesis Proposal

Tuesday, December 4, 2018 - 11:00am to 12:00pm

Location:

Traffic21 Classroom 6501 Gates Hillman Centers

Speaker:

SAURABH KADEKODI, Ph.D. Student http://www.cs.cmu.edu/~saukad/

Exploiting heterogeneity in cluster storage systems


Speaker: Saurabh Kadekodi

Location GHC 6501


Exploiting heterogeneity in cluster storage systems

Large-scale storage systems include a heterogeneous mix of storage devices with different reliability, capacity and performance characteristics. This dissertation explores the benefits of explicitly factoring in heterogeneity in the characteristics of storage devices belonging to a single tier, for cheaper redundancy and more load-balanced data placement.

Data reliability in today's cluster storage systems is agnostic to the failure rate differences observed between storage devices in the same tier. Using a production dataset of over 100,000 hard disks, we show that there is a significant diversity in hard disk failure rates, as a function of make/model. We build a case for failure rate tailored redundancy, and design and prototype HeART (Heterogeneity-Aware Redundancy Tuner), an online automated framework to aid the process. HeART continuously monitors the failure rates of different disk groups, and employs robust online algorithms to safely and accurately identify redundancy schemes for each disk group, while meeting a defined reliability target. This leads to lower level of redundancy for devices with low failure rate and more redundancy for devices with high failure rate. Redundancy informed by HeART would result in 13-44% fewer disks than one-scheme-fits-all approaches in the production dataset: 33-44% compared to 3-replication and 13-19% compared to erasure codes like 6-of-9 or 10-of-14.

Technological advancements make capacity heterogeneity also prevalent, and unavoidable in large cluster storage systems. Data placement algorithms agnostic to capacity heterogeneity often cause heavy spindle imbalance across the disk fleet. This load imbalance worsens when there is data heat (accesses per unit time) heterogeneity across different clients. We propose a data placement algorithm that accounts for each dataset's current heat and expected rate of cooling in addition to storage device capacity heterogeneity. These simple heuristics allow for data placement that results in higher storage device utilization, lower rebalancing cost and lower tail latencies.

Thesis Committee:
Gregory R. Ganger (Co-chair)
Rashmi Vinayak (Co-chair)
Garth Gibson
Remzi Arpaci-Dusseau (University of Wisconsin-Madison)

Copy of Thesis Summary

For More Information, Contact:

Keywords:

Thesis Proposal