Grant: Research Tools and Instruments

The requested infrastructure is a disaggregated and heterogeneous computing platform (DHCP) to support research into next generation distributed/parallel (scale-out) data management systems. This platform includes independent pools of compute nodes, memory nodes, storage nodes, and hardware accelerator (GPU, FPGA) nodes that are physically separated but are connected via an ultra-fast, low-latency network. DHCP enables all resources on the network to be shared. Compute nodes contain a small amount of memory to serve as data buffer, and the memory nodes and storage nodes may have light computing capability to allow sharing. The compute capability on the storage and memory nodes also facilitates near-data processing to avoid excessive data shipping to compute nodes. An important development of the past decade is the introduction of cloud computing platforms. According to a recent study, over 50% of all deployments are now on the cloud and the expectation is that over 60% will be on the cloud in the next 18 months. The advantages of moving to the cloud are well-known, and a strong case can be made that it is the most suitable platform for many computing tasks. Current cloud data centres typically use a rack-and-blade design where each blade contains some compute elements, some memory, usually some storage, and perhaps some accelerators tightly connected and multiple blades are mounted on a rack. This converged data centre architecture is now starting to experience difficulties in adequately serving emerging big data applications with highly varying application requirements. An important expectation of cloud computing is to provide elasticity to accommodate these extreme application needs only when they arise. However, converged data centre architecture is stressed in providing this elasticity and usually over-provisions (allocates from the beginning the maximum resources that may be needed at some point) resulting in low resource utilization. For example, Google cluster reports on the average 40% utilization while Alibaba cluster reports 20-40% utilization. Over-provisioned memory also causes what is called stranded memory where preallocated memory goes unused. The ultimate result is the waste of resources that is costly - server memory constitutes 50% of the server cost on Microsoft Azure. DHCPs are now being investigated as a much preferable way to fulfill the promises of cloud platforms. They allow better sharing of resources that results in reduced cost. The ability to carry out this research requires a DHCP infrastructure. The goal of the proposed infrastructure is to facilitate research into the design of big data management systems, tools, and techniques on top a DHCP and to investigate the design tradeoffs. The focus of the research is on the data management layer, but attention will also be paid to the underlying systems issues.