Title: Parallel Hash Table Implementation and Performance on 64-core Tilera Chips For Network Speaker: Bob Benner, Sandia National Laboratories Date/Time: Thursday, August 27, 2009, 3:00-4:00 pm Location: CSRI Building, Room 90 (Sandia NM) Brief Abstract: Hash tables are an important means of storing data when analyzing network traffic and in other computer security applications. Parallel hash tables have a long, colorful, and often dismal history, dating back to the earliest commercial shared and distributed memory systems of the mid-1980’s. We have recently implemented several hash algorithms on the 64-core Tilera chip and have some very promising results. On the 2nd generation, Tile64Pro chip we see parallel efficiencies of 55-88% for various hash algorithms and hash table sizes on up to 57 tile processors. Why 57 tiles? The current default system configuration reserves 7 tiles for various system tasks, particularly for driving high speed interfaces (1 gigE, 10 gigE). Several other architectural factors must be accounted for when considering potential applications for Tilera. For example, the 64 cores are integer cores – all floating point computations are done via emulation. This is ideal for network traffic analysis – the only floating point we do is timing statistics for performance analysis. Most important, the chip supports both shared memory and private memory – but, since both kinds of memory are off-chip, shared memory implementations are preferable. The chip also has five independent interprocessor communication networks, some of which are reserved for system tasks. And I won’t even attempt to describe user-defined and system-controlled cache management strategies until the talk itself. My hope is that this talk will stimulate your thinking beyond the realm of hash tables and network traffic analysis - on how to apply emerging multicore chips and GPU’s to a wide spectrum of numerical algorithms, methods and applications, i.e. those that are of personal interest to you.CSRI POC: Zhaofang Wen, 284-0206 |