Mahbod Afarin

JetStream (MICRO'21): Event-Driven Hardware Accelerator for Streaming Graph Analytics

Jun' 20 - Apr' 21

University of California, Riverside, California, USA

• First Streaming Graph Accelerator: Developed JetStream, the first accelerator to support operations on streaming graphs (or dynamic graphs).
• New Asynchronous Streaming Algorithms: JetStream subsumes the capabilities of GraphBolt and KickStarter software frameworks that allow edge deletions.
• Large Performance Improvements that improve with smaller batch sizes: JetStream substantially outperforms both software frameworks. In addition, its advantage grows as the batch size is reduced, making it suitable for near real-time analytics.

MEGA (MICRO'23): MEGA Evolving Graph Accelerator

Jun' 22 - May' 23

University of California, Riverside, California, USA

• First Evolving Graph Accelerator: MEGA, the first accelerator for evolving graph workloads, provides support for multiple snapshots executing simultaneously.
• Batch-Oriented Execution:: Batch-oriented execution exploits the similarity of the graph across snapshots to reuse similar edge-fetches and minimize redundant execution of batches.
• Batch Pipelining:: We explore optimizations to the workflow to improve concurrency such as allowing multiple concurrent batches, and using pipelining across batches.

CommonGraph (ASPLOS'23): Graph Analytics on Evolving Data

Aug' 21 - Jun' 22

University of California, Riverside, California, USA

• Converting Expensive Deletions to Additions: Developed a new approach to evolving graphs analysis that avoids processing of deletions by converting them to additions to the CommonGraph. CommonGraph facilitates work sharing and parallelism across snapshots.
• Triangular Grid Data Structure for Direct-Hop and Work-Sharing Algorithm:
• Developed a new structure called Triangular Grid that exposes work-sharing opportunities across snapshots. A Steiner Tree formulation finds a solution that maximizes work-sharing.
• Designed a graph representation that avoids the need to mutate graphs and enables reuse of edges by snapshots that share them.
• Implemented the CommonGraph that exploits the above ideas and delivers considerable speedups over KickStarter.

Core Graph (EuroSys'24): Exploiting Edge Centrality to Speedup the Evaluation of Iterative Graph Queries

May' 21 - May' 23

University of California, Riverside, California, USA

• Core Graph Identification and Exploitation:
• Our study of multiple kinds of graph queries on irregular graphs shows that the solution of a query is determined by a small fraction of total edges, i.e. critical edges. Also, many edges recur frequently across critical edge sets of different queries.
• Present algorithms for finding a core graph by solving a small set of queries to identify most non-zero centrality edges.
• Exploit core graph and present a new optimization that improves the efficiency of the 2phase evaluation while producing 100\% precise results.
• Experimental Results: For the 2.586 billion Friendster graph, across six kinds of queries, our approach yielded core graphs containing 5.42% to 10.45% edges and precise results for 97.1 - 99.9% vertices.

Doctor of Philosophy (Ph.D.), Computer Science

Jan' 20 - Jun' 24 (Expected)

University of California, Riverside, California, USA

Thesis: "Hardware-Software Approaches for Accelerating Graph Processing Workloads"

Advisors: Professor Rajiv Gupta & Professor Nael Abu-Ghazaleh

GPA: 3.86/4

Master of Science (M.Sc.), Computer Engineering (Computer System Architecture)

Sep' 15 - Jan' 18

Sharif University of Technology, Tehran, Iran

Thesis: "Improving Manufacturing Yield and Life Cycle of Special Purpose SIMT Processors for Inexact Computing" (Thesis Grade: Excellent)

Advisors: Professor Shaahin Hessab

GPA: 4/4 (19.03/20) - Ranked 7th in terms of total GPA among 83 Computer Engineering students (Top 8%)

Bachelor of Science (B.Sc.), Computer Engineering (Computer System Architecture)

Sep' 11 - Jun' 15

Shahed University, Tehran, Iran

Thesis: "Comparing different types of software for designing with SystemC and implementing Mano processor with SystemC" (Thesis Grade: Excellent)

Advisors: Professor Naser Mohammadzadeh

GPA: 3.63/4 (17.53/20) - Ranked 1st among all Computer Engineering graduate students