Power Aware Dynamic Provisioning of HPC Networks

Abstract

Future exascale systems are under increased pressure to find power savings. The network, while it consumes a considerable amount of power is often left out of the picture when discussing total system power. Even when network power is being considered, the references are frequently a decade or older and rely on models that lack validation on modern interconnects. In this work we explore how dynamic mechanisms of an Infiniband network save power and at what granularity we can engage these features. We explore this within the context of the host controller adapter (HCA) on the node and for the fabric, i.e. switches, using three different mechanisms of dynamic link width, frequency and disabling of links for QLogic and Mellanox systems. Our results show that while there is some potential for modest power savings, real world systems need to improved responsiveness to adjustments in order to fully leverage these savings.

A survey on software methods to improve the energy efficiency of parallel computing

“…Improving communication EE consists of three groups of methods: (a) saving network fabric power with power-aware networks (Alonso et al, 2006; Conner et al, 2007; Groves and Grant, 2015; Miwa et al, 2014; Miwa and Nakamura, 2015; Nedevschi et al, 2008; Saravanan et al, 2013, 2014; Totoni et al, 2014a), (b) improving communication EE using energy-aware data transfer algorithms (Alan et al, 2015; Jana and Chapman 2015; Jana et al, 2014a, 2014b, 2014c; Kandalla et al, 2010; Lim et al, 2006; Venkatesh et al, 2015; Vishnu et al, 2013), and (c) decreasing network traffic using communication-avoiding algorithms (Demmel et al, 2013; Grigori et al, 2011).…”

“…Some of these methods depend on energy improvement of compiler techniques (Hsu and Kremer, 2003b;Keramidas et al, 2010;Leon et al, 2015;Saputra et al, 2002). (3) Improving communication EE consists of three groups of methods: (a) saving network fabric power with power-aware networks (Alonso et al, 2006;Conner et al, 2007;Groves and Grant, 2015;Miwa et al, 2014;Miwa and Nakamura, 2015;Nedevschi et al, 2008;Saravanan et al, 2013Saravanan et al, , 2014Totoni et al, 2014a), (b) improving communication EE using energy-aware data transfer algorithms (Alan et al, 2015; Jana and Chapman 2015; Jana et al, 2014aJana et al, , 2014bJana et al, , 2014cKandalla et al, 2010;Lim et al, 2006;Venkatesh et al, 2015;Vishnu et al, 2013), and (c) decreasing network traffic using communication-avoiding algorithms (Demmel et al, 2013;Grigori et al, 2011). (4) Automatic tuning: This category treats a parallel application as a black box and improves its EE by tuning several energy and performance parameters (Balaprakash et al, 2013;Gschwandtner et al, 2014;Jordan et al, 2012;Miceli et al, 2012;Rahman et al, 2011;Tiwari et al, 2012).…”

“…At exascale, the network can consume 10–20% of the total system power (Groves and Grant, 2015). Some experts estimate that future networks may account for 30% of the total supercomputer power budgets (Dickov et al, 2014).…”

“…Groves and Grant (2015) investigate existing InfiniBand products in terms of power saving. As reported, InfiniBand’s WRPS provides potential for modest power savings.…”

“…Some experts estimate that future networks may account for 30% of the total supercomputer power budgets (Dickov et al, 2014). In order to improve the EE of the interconnection network, many innovative technologies have been developed, including on-die interconnect fabric, interchip network integration, energy proportional networks, and power-aware networks (Alonso et al, 2006;Groves and Grant, 2015;Miwa et al, 2014;Miwa and Nakamura, 2015;Nedevschi et al, 2008;Saravanan et al, 2013Saravanan et al, , 2014. Power-aware networks can save power consumed by the network fabric using techniques such as dynamic link width and frequency, and on/off links.…”