Title: Adding machine intelligence to hybrid memory management

Thaleia Dimitra Doudali

School of Computer Science

College of Computing

Georgia Institute of Technology

https://www.cc.gatech.edu/~tdoudali/

Date: Tuesday December 15th 2020

Time: 2:00 PM - 4:00 PM (EST)

Location: https://bluejeans.com/8193487972

Committee:

Dr. Ada Gavrilovska (Advisor, School of Computer Science, Georgia Institute of Technology)

Dr. Vivek Sarkar (School of Computer Science, Georgia Institute of Technology)

Dr. Alexey Tumanov (School of Computer Science, Georgia Institute of Technology)

Dr. Tushar Krishna (School of Electrical and Computer Engineering, Georgia Institute of Technology)

Dr. Sudhanva Gurumurthi (Principal Member of the Technical Staff, AMD)

Abstract:

The integration of non volatile memory technologies to the main memory substrate enables massive capacities at a reasonable cost in return for slower access speeds. Preserving the application performance levels of traditional homogeneous memories is feasible with fine-tuned data management approaches. At the same time, resource management solutions augmented with machine learning show great promise for fine-tuning system configuration knobs and predicting future behaviors. This thesis explores such trends, reveals new insights and closes the application performance gap left by existing solutions, with the following contributions.

First, this thesis builds Kleio; a hybrid memory page scheduler with machine intelligence. Kleio deploys Recurrent Neural Networks to learn memory access patterns and optimizes the selection of which data to dynamically move across the memory units. Kleio cleverly focuses the machine learning on the page subset whose timely movement will reveal most application performance improvement, while preserving history-based lightweight management for the rest of the pages. In this way, Kleio bridges on average 80% of the relative existing performance gap, while laying the grounds for practical machine intelligent data management with manageable learning overheads.

Second, this thesis contributes Cori; a system-level solution for tuning the operational frequency of periodic page schedulers for hybrid memories. Cori synthesizes information on data reuse to properly identify the data movement frequencies to be tested, reducing by 5x the number of tuning trials compared to existing empirical or insight-less tuning approaches. In this way, Cori delivers application performance levels within 3% from the case of optimally selected frequency, eliminating the 10%-100% performance gap created when using frequencies currently adopted by related works. Such improvements translate to substantial gains when considering the scale of emerging supercomputing and datacenter systems. 

Finally, this thesis proposes a holistic machine intelligent page scheduling solution that brings together insights from Cori and the design principles of Kleio. The outcome is a page scheduler that leverages similarity in page access patterns to afford use of machine intelligent management on more pages than Kleio, while amortizing training overheads. Combining this with the use of a cost-efficient operational frequency, as identified by Cori, enables maximum performance improvements that otherwise cannot be practically realized with acceptable costs.

Additional Meeting Details:

Meeting URL

https://bluejeans.com/8193487972?src=join_info

Meeting ID

819 348 797 2

Want to dial in from a phone?

Dial one of the following numbers:

+1.408.419.1715 (United States(San Jose))

+1.408.915.6290 (United States(San Jose))

(see all numbers - https://www.bluejeans.com/numbers)

Enter the meeting ID and passcode followed by #

Connecting from a room system?

Dial: bjn.vc or 199.48.152.152 and enter your meeting ID & passcode