Virtualization in the age of heterogeneous machines

Since their invention over 40 years ago, virtual machines have been used to virtualize one or more von Neumann processors and their associated peripherals. System virtual machines provide the illusion that the user has their own instance of a physical machine with a given instruction set architecture (ISA). Process virtual machines provide the illusion of running on a synthetic architecture independent of the underlying ISA, generally for the purpose of supporting a high-level language. To continue the historical trend of exponential increase in computational power in the face of limits on clock frequency scaling, we must find ways to harness the inherent parallelism of billions of transistors. I contend that multi-core chips are a fatally flawed approach - instead, maximum performance will be achieved by using heterogeneous chips and systems that combine customized and customizable computational substrates that achieve very high performance by closely matching the computational and communications structures of the application at hand. Such chips might look like a mashup of a conventional multicore, a GPU, an FPGA, some ASICs, and a DSP. But programming them with current technologies would be nightmarishly complex, portability would be lost, and innovation between chip generations would be severely limited. The answer (of course) is virtualization, and at both the device level and the language level. In this talk I will illustrate some challenges and potential solutions in the context of IBM's Liquid Metal project, in which we are designing a new high-level language (Lime) and compiler/runtime technology to virtualize the underlying computational devices by providing a uniform semantic model. I will also discuss problems (and opportunities) that this raises at the operating system and data center levels, particularly with computational elements like FPGAs for which "context switching" is currently either extremely expensive or simply impossible.