Jekyll & Hyde: a heterogeneous key-value store.
Event details
| Date | 19.06.2012 |
| Hour | 10:00 › 11:00 |
| Speaker | Dr. Dushyanth Narayanan, Microsoft Research Cambridge |
| Location | |
| Category | Conferences - Seminars |
Abstract:
Data center hardware is about to get very heterogeneous in terms of CPUs, interconnects, and memory because of developments like ARM-based microservers, custom network fabrics, and hybrid memory cubes all on the horizon. What does this imply for systems design especially for cloud infrastructures?
In this talk I will describe our recent work on “Jekyll and Hyde”: an in-memory key-value store for asymmetric clusters. The store consists of two in-memory replicas: Jekyll and Hyde. “Jekyll” is a single big memory server (192 GB in our prototype) that holds one of the replicas. “Hyde” is a larger number of lower-power machines across which the second replica is sharded. In our prototype we use a cluster of older-generation Intel processors but these could also be ARM or Atom-based microservers. By directing transactions and random-access queries to the Jekyll and parallelizable scan-based analytics to the Hyde, we leverage the strengths of each. The transactional workload benefits from the cache coherence and fast interconnect of the Jekyll, while the analytics benefit from the high aggregate bandwidth of the Hyde. An additional benefit is that the scan workloads do not compete for memory bandwidth or pollute the cache of the transactional workloads.
If time permits I’ll then talk briefly about our ongoing work on locality-preserving partitioning.
Bio:
Dushyanth Narayanan is a researcher in the Systems and Networking group at Microsoft Research Cambridge. He has worked on mobile computing, storage, energy-efficiency, and most recently in data center and cloud architectures.
Data center hardware is about to get very heterogeneous in terms of CPUs, interconnects, and memory because of developments like ARM-based microservers, custom network fabrics, and hybrid memory cubes all on the horizon. What does this imply for systems design especially for cloud infrastructures?
In this talk I will describe our recent work on “Jekyll and Hyde”: an in-memory key-value store for asymmetric clusters. The store consists of two in-memory replicas: Jekyll and Hyde. “Jekyll” is a single big memory server (192 GB in our prototype) that holds one of the replicas. “Hyde” is a larger number of lower-power machines across which the second replica is sharded. In our prototype we use a cluster of older-generation Intel processors but these could also be ARM or Atom-based microservers. By directing transactions and random-access queries to the Jekyll and parallelizable scan-based analytics to the Hyde, we leverage the strengths of each. The transactional workload benefits from the cache coherence and fast interconnect of the Jekyll, while the analytics benefit from the high aggregate bandwidth of the Hyde. An additional benefit is that the scan workloads do not compete for memory bandwidth or pollute the cache of the transactional workloads.
If time permits I’ll then talk briefly about our ongoing work on locality-preserving partitioning.
Bio:
Dushyanth Narayanan is a researcher in the Systems and Networking group at Microsoft Research Cambridge. He has worked on mobile computing, storage, energy-efficiency, and most recently in data center and cloud architectures.
Links
Practical information
- General public
- Free
Organizer
- SuRI 2012
Contact
- Simone Muller