Software-Managed Power Reduction in Infiniband Links

Abstract

Abstract-Interconnection network represents the backbone of large-scale supercomputers. While the computation has become more energy proportional, the interconnection network has started to become one of the important consumers in the system. Still, interconnection network has received little attention in terms of efficient power consumption. Several hardware-based schemes have been proposed that function either by turning off the links or lowering the voltage/frequency in links when idle or with low usage. Although effective in certain cases, these schemes share the common drawback that they do not have enough global information about the application to manage network power efficiently. This paper, instead, proposes a software prediction support for shutting down inactive lanes of an Infiniband (IB) link.Using n-gram extraction technique as a core of the prediction algorithm we are able to predict with high accuracy when links will be unused. The prediction algorithm detects the repeating patterns of MPI calls during the program execution, on which basis the prediction of the following MPI calls can be realized with relatively high precision. This allow us to know on time when links will be active/idle and invoke the explicit turn-off/on lanes calls. Our power saving mechanism is adopted to be executed within MPI allowing a large number of existing MPI programs to be run without any source code modification. Results show an average power savings in IB switches up to around 33% over the representative HPC applications. On the other hand, HPC applications average suffer only up to 1% increase in execution time.

Power Aware Dynamic Provisioning of HPC Networks

“…In summary, current network fabric has much too long a delay to apply the rapid adjustments to the network that we see in more recent work [4]. While new technology may change this, we have yet to evaluate the latest offerings from vendors.…”

“…Dickov et al [4] simulated predictively reducing the link width of the network fabric by annotating the MPI layer. Their work references a Mellanox whitepaper [1], which claims a reduction of switch (SX6036) power to 43% of the nominal power, by reducing all links to 1X width.…”

PerfBound: Conserving Energy with Bounded Overheads in On/Off-Based HPC Interconnects

“…In their approach, during periods of inactivity, link rates are reduced to a lower link bandwidth to save energy. B. Dickov, et al, [27] present power savings for InfiniBand networks. They show similar energy saving potential is available in HPC systems and use prediction based methods to save link energy.…”

Prediction of the impact of network switch utilization on application performance via active measurement

“…Other work has been more focused on exploring the usefulness of new network topologies to achieve significant reductions in both network cost and network power, while still providing a balance of high global and high local bandwidth [41]. Previous work [33,12] considered how power may reduce interconnection networks' capacity. There have been several prior efforts to reduce power in interconnection networks with a particular focus on reducing power on infrequently used links.…”