Studies in Exascale Computer Architecture: Interconnect, Resiliency, and Checkpointing

exascale supercomputer architecture;kilo-core on-chip interconnect topology;checkpoint/restart fault tolerance;Computer Science;Engineering;Computer Science & Engineering

Abeyratne, SandunmaleeDas, Reetuparna ;

University of Michigan

关键词: exascale supercomputer architecture;    kilo-core on-chip interconnect topology;    checkpoint/restart fault tolerance;    Computer Science;    Engineering;    Computer Science & Engineering;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/137096/sabeyrat_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Today’s supercomputers are built from the state-of-the-art components to extract as much performance as possible to solve the most computationally intensive problems in the world. Building the next generation of exascale supercomputers, however, would require re-architecting many of these components to extract over 50x more performance than the current fastest supercomputer in the United States. To contribute towards this goal, two aspects of the compute node architecture were examined in this thesis: the on-chip interconnect topology and the memory and storage checkpointing platforms.As a first step, a skeleton exascale system was modeled to meet 1 exaflop of performance along with 100 petabytes of main memory. The model revealed that large kilo-core processors would be necessary to meet the exaflop performance goal; existing topologies, however, would not scale to those levels. To address this new challenge, we investigated and proposed asymmetric high-radix topologies that decoupled local and global communications and used different radix routers for switching network traffic at each level. The proposed topologies scaled more readily to higher numbers of cores with better latency and energy consumption than before.The vast number of components that the model revealed would be needed in these exascale systems cautioned towards better fault tolerance mechanisms. To address this challenge, we showed that local checkpoints within the compute node can be saved to a hybrid DRAM and SSD platform in order to write them faster without wearing out the SSD or consuming a lot of energy. A hybrid checkpointing platform allowed more frequent checkpoints to be made without sacrificing performance. Subsequently, we proposed switching to a DIMM-based SSD in order to perform fine-grained I/O operations that would be integral in interleaving checkpointing and computation while still providing persistence guarantees. Two more techniques that consolidate and overlap checkpointing were designed to better hide the checkpointing latency to the SSD.