Keynote talks

Current trends in computer architectures: Multi-cores, Many-cores and Special-cores

Power thermal and process limitations encourage modern processors to integrate few cores on the same die in order to maintain overall performance growth, expected by the industry. While two years ago, most processors where single core configuration, the majority of the current processors contain dual or quad cores and the number of cores on die is expected to grow over time.

One can observe three different trends in the development of future computer architectures: (1) Multi-cores - integrating a set of cores, each of them preserves the same "single thread" performance as previous generation, (2) Many-cores - integrating large number of cores, trading single threaded performance with MTL (multi-threaded level) performance and (3) special cores - an integration of multi-cores with many cores, fixed function logic and/or programmable logic such as FPGA.

The success of each of these trends depends on both hardware and software technologies. In my talk I will explain why the market is moving from single core to multi/many/special core architectures, examine each of the new trends and try to predict the success/failure factor (software and hardware) for each of them.

Harnessing Massive Parallelism in the era of Parallelism for the Masses

As we embrace this new era of multicore and heterogeneous processors, we find ourselves ill-equipped to leverage the full potential of increasing parallelism. There are no widely-adopted parallel programming languages. With massive datasets, the growing importance of sensor-based systems, and increasingly large data models the I/O wall is looking more daunting than the memory wall. Current high-end software technology is not easily used by non-sophisticated developers.

However, the stakes are high for the entire industry: we must effectively leverage parallel systems, even down to laptops and embedded systems, to profit from the multicore revolution. In response, there is a renewed focus on creatively addressing parallelism and the mismatch between "free" compute versus expensive communication. Parallel languages are receiving more serious attention again. Researchers are investigating new programming models such as transactional memory. There is a renewed focus on parallel compiler functionality and performance. Performance tools for parallel programs are improving. As a result, the effort needed to create efficient, scalable programs should noticeably improve over the next few years.

On the very high end, minimization of noise and communication delay remain paramount. Operating systems must be as simple and predictable as possible to minimize jitter and to increase reliability. Synchronization and collective communication must be efficiently supported. Whenever possible, we must find ways of bringing the program to the data instead of the other way around. I use examples from IBM Blue Gene systems to illustrate some high-end scaling strategies for systems software and for application porting and tuning.