Better GPU Hash Tables

Hello CUDA community,

We're happy to announce our next online meetup!
On October 12th (9 am (PDT), 19:00 (MSK - UTC+3)), we'll talk about GPU hash tables. We will add a link to the Zoom meeting closer to the event so stay tuned.

Abstract:
Building an efficient static GPU hash table is influenced by four major design decisions: probing scheme, bucket size, probe complexity, and placement strategy. These decisions affect the number of probes an insertion or query operation performs, which is the most important factor influencing the performance of these operations. In this talk, I will discuss the lessons learned in designing and implementing three different static GPU hash tables. We will discuss the four choices affecting hash table design and their associated tradeoffs. I will show that a bucketed cuckoo hash table (BCHT) outperforms other alternatives. BCHT achieves load factors as high as 99% and an average probe count of only 1.43, 1.39, and 2.8 during insertion, positive and negative queries, respectively.

Speaker:
Muhammad Awad is a Ph.D. candidate in the Electrical and Computer Engineering Department at UC Davis. Working in John Owens's research group, he designs and builds data structures targeting the GPU. His recent work includes developing a dynamic graph data structure, dynamic B-Tree, and static hash tables.

Agenda:

• Welcome message
• Raffle of free CLion licenses
• The talk
• Post meetup discussions