Disaggregating Non-Volatile Memory for Throughput-Oriented Genomics Workloads

Abstract

Massive exploitation of next-generation sequencing technologies requires dealing with both: huge amounts of data and complex bioinformatics pipelines. Computing architectures have evolved to deal with these problems, enabling approaches that were unfeasible years ago: accelerators and Non-Volatile Memories (NVM) are becoming widely used to enhance the most demanding workloads. However, bioinformatics workloads are usually part of bigger pipelines with different and dynamic needs in terms of resources. The introduction of Software Defined Infrastructures (SDI) for data centers provides roots to dramatically increase the efficiency in the management of infrastructures. SDI enables new ways to structure hardware resources through disaggregation, and provides new hardware composability and sharing mechanisms to deploy workloads in more flexible ways. In this paper we study a state-of-theart genomics application, SMUFIN, aiming to address the challenges of future HPC facilities.

Leveraging disaggregated accelerators and non-volatile memories to improve the efficiency of modern datacenters

“…This behavior can be observed in our previous experiments [9]. Consequently, the idea behind the first proposed policy is, when unused resources must be found and allocated to workloads, to compose many resources into a single one, over-provisioning, as long as the workload will benefit from it.…”

“…Bandwidth-bound: represents workloads that are sensitive to bandwidth, so multiple concurrent workloads running in the same device may impact performance if bandwidth capacity is exceeded. We use our work on SMUFIN [58][7], to model the behavior of this workload and the results presented in [9]. This model is then fed to the placement policies to assess how to allocate resources.…”

“…Bandwidth-bound: represents workloads that are sensitive to bandwidth, and so multiple concurrent workloads running in the same the device may have an impact on performance if bandwidth capacity is exceeded. We use our work on SMUFIN [58][7], to model the behavior of this workload and the results presented in [9], summarized in figure 4 6 concurrent executions were sharing the same composition. This model shows how over-provisioning the workload may have some benefits both in terms of performance as well as workload's concurrence threshold increase without experiencing degradation.…”

A Case For Intra-rack Resource Disaggregation in HPC

“…This is in contrast to many systems today where nodes are allocated to applications as a unit with identical ixed-sized resources; any resources inside nodes that the application does not use have no choice but to idle. Following the trend for hardware specialization and the desire to better utilize resources as systems scale up, resource disaggregation across the system or a group of racks has been actively researched and deployed in commercial hyperscale datacenters in Google, Facebook, and others [17,55,64]. In addition, many studies focus on disaggregation of GPUs [35] and memory capacity [33,67].…”

“…For this reason, a few studies argue for rack-level disaggregation of GPUs [34,75]. In addition, other studies rely on electrical networks [31] in some cases with the aid of software deined networks (SDNs) [17]. Finally, how the software stack should adapt to best take advantage of resource disaggregation remains a signiicant challenge both for the OS [71] and job scheduler [10,27].…”