Position: Postdoctoral Researcher
Current Institution: VMware Research
Abstract: Observably Consistent and Performant Distributed Systems
The tradeoff between performance and correctness is pervasive across computer systems including shared-memory multiprocessors databases and local file systems. The same tradeoff exists in distributed storage systems as well; designers must often choose consistency or performance but not both. In my research I design and build distributed storage systems that provide strong guarantees yet also perform well. These new systems are designed based on a central idea: a system can remain inconsistent as long as external entities do not observe the system state and consistency must be enforced only when the state is externalized to the outside world. This idea enables a system to defer and group expensive work thereby improving performance. Based on this insight I design two novel distributed storage systems. First I propose consistency-aware durability or CAD a new approach to durability in distributed storage. The key idea behind CAD is to shift the point of durability from writes to reads. By delaying durability upon writes CAD enables high performance; however by ensuring the durability of data before serving reads CAD enables the construction of stronger consistency models. Second I build NilPaxos a new replication protocol that improves the performance of strongly consistent storage. Standard approaches to building strongly consistent storage incur high coordination overhead. The main insight in NilPaxos is that much of this coordination is unnecessary if the storage-interface semantic is carefully exploited. Particularly my work realizes that many update interfaces in modern key-value stores do not expose system state; we call such interfaces nil-externalizing interfaces. NilPaxos takes advantage of nil-external updates by ordering and executing them lazily reducing coordination overhead thus improving performance. NilPaxos enforces consistency before state is externalized through subsequent operations.
Bio:
Aishwarya Ganesan is a postdoctoral researcher at VMware Research. She recently earned her PhD from the University of Wisconsin-Madison. She is broadly interested in distributed systems and storage systems. Her research addresses the fundamental tradeoff between strong consistency and performance in distributed storage systems. Her work on distributed storage reliability has exposed many severe bugs in popular distributed systems. Ideas from her research on corruption-tolerant replication are implemented in a financial database. Her work has appeared in top systems venues such as OSDI SOSP and FAST. Her research has been recognized with best-paper awards at FAST 20 and FAST 18 and a best paper award nomination at FAST 17. She is a recipient of a Facebook 2019 PhD Fellowship. She also received the graduate student instructor award for teaching graduate-level distributed systems at UW Madison.