I am a fourth year PhD student in the CSE department at the University of Michigan, Ann Arbor.  I work with Thomas Wenisch in the Advanced Computer Architecture Lab (ACAL).  My current work focuses on large scale, energy efficient computing.

I received my BS in Electrical and Computer Engineering from Cornell University in Ithaca, NY.

Research Interests:

I am interested in investigating how best to use nonvolatile memory (NVRAM), such as phase change, memristor, and SST, to reduce the complexity and improve the performance of database applications. Additionally, an NVRAM memory hierarchy will likely require a cache and DRAM write buffer. I am investigating a cache hierarchy that support ACID transactions (such as in database workloads) while providing the performance we have come to expect from deep memory hierarchies.

Previous Projects

Flash Database Query Optimization

Databases must make many different decisions regarding how to process queries, including choice of index scans vs relational scans, join algorithms, and join order. Which decision is best depends on the data involved as well as the characteristics of the device where that data is stored. Based on previous models and rules of thumb we believe that the optimal choices for databases will shift for a large number of queries when moving from disk to flash, resulting in a large performance gain (relative to the incorrect query plan calculated for disk).

While the intuition seemed compelling, we found that this is largely not the case. Typical database set ups contain a large number of tuples per page. As data is frequently randomly distributed you will expect to retrieve a large number of pages for even a small selectivity of rows. We observe that the range of selectivities where scan decision differs between disk and flash is so small that it represents a negligible fraction of queries and that the price for an incorrect decision is small. Additionally, existing join algorithms perform so similarly on Flash SSDs that the decision is largely irrelevant. We have yet to investigate research algorithms for scans and joins, and believe this will be prohibitively difficult.

Steven Pelley, Kristen LeFevre, and Thomas F. Wenisch
"Do Query Optimizers Need to be SSD-aware?"
VLDB 2011: 2nd International Workshop on Accelerating Data Management Systems Using Modern Processor and Storage Systems. Seattle, WA. Sept. 2011.

Power Routing

Power infrastructure is one of the most expensive components of modern data centers. The need for high availability, which is commonly achieved through redundancy, compounds these costs. Optimistic schemes such as power capping allow the data center to provide less power capacity with the understanding that the data center is rarely fully utilized. Power Routing takes this a step further by exploiting the redundancy mechanism in Power Distribution Units (PDUs) to remove local power constraints across the data center, thus reducing capital expenses.

Steven Pelley, David Meisner, Pooya Zandevakili, Thomas F. Wenisch, and Jack Underwood
Power Routing: Dynamic Power Provisioning in the Data Center
ASPLOS 2010. Pittsburgh, Pennsylvania

Maelstrom:

Much of the inefficiency of data centers stems from the difficulty of removing heat from the computer room.  Due to the complicated airflows typically found in data centers, we believe that a full Computational Fluid Dynamics (CFD) analysis is required to accurately predict airflows.  Such predictions can be used to intelligently schedule tasks throughout the physical data center space, as well as analyse data center operation offline.  We are working on near-real-time CFD for the data center.

Project discontinued. Server power efficiency and cooling infrastructure provide much more opportunity to reduce cooling costs without the need to interfere with workloads. As baseline cooling efficiency increases the opportunity for active, intelligent cooling diminishes.