Search NASA⌕ Search

Engineering topics

Chong, F T

Publications and source records attributed to Chong, F T.

Exploiting Data Similarity to Reduce Memory Footprints

Memory size has long limited large-scale applications on high-performance computing (HPC) systems. Since compute nodes frequently do not have swap space, physical memory often limits problem sizes. Increasing core counts per chip and power density constraints, which limit the number of DIMMs per node, have exacerbated this problem. Further, DRAM constitutes a significant portion of overall HPC system cost. Therefore, instead of adding more DRAM to the nodes, mechanisms to manage memory usage more efficiently - preferably transparently - could increase effective DRAM capacity and thus the benefit of multicore nodes for HPC systems. MPI application processes often exhibit significant data similarity. These data regions occupy multiple physical locations across the individual rank processes within a multicore node and thus offer a potential savings in memory capacity. These regions, primarily residing in heap, are dynamic, which makes them difficult to manage statically. Our novel memory allocation library, SBLLmalloc, automatically identifies identical memory blocks and merges them into a single copy. SBLLmalloc does not require application or OS changes since we implement it as a user-level library. Overall, we demonstrate that SBLLmalloc reduces the memory footprint of a range of MPI applications by 32.03% on average and up to 60.87%. Further, SBLLmalloc supports problem sizes for IRS over 21.36% larger than using standard memory management techniques, thus significantly increasing effective system size. Similarly, SBLLmalloc requires 43.75% fewer nodes than standard memory management techniques to solve an AMG problem.

97 MATHEMATICS AND COMPUTING↗

Fighting Fire with Fire: Modeling the Datacenter-Scale Effects of Targeted Superlattice Thermal Management

Local thermal hot-spots in microprocessors lead to worst case provisioning of global cooling resources, especially in large-scale systems. However, efficiency of cooling solutions degrade non-linearly with supply temperature, resulting in high power consumption and cost in cooling - 50 {approx} 100% of IT power. Recent advances in active cooling techniques have shown on-chip thermoelectric coolers (TECs) to be very efficient at selectively eliminating small hot-spots, where applying current to a superlattice film deposited between silicon and the heat spreader results in a Peltier effect that spreads the heat and lowers the temperature of the hot-spot significantly to improve chip reliability. In this paper, we propose that hot-spot mitigation using thermoelectric coolers can be used as a power management mechanism to allow global coolers to be provisioned for a better worst case temperature leading to substantial savings in cooling power. In order to quantify the potential power savings from using TECs in data center servers, we present a detailed power model that integrates on-chip dynamic and leakage power sources, heat diffusion through the entire chip, TEC and global cooler efficiencies, and all their mutual interactions. Our multiscale analysis shows that, for a typical data center, TECs allow global coolers to operate at higher temperatures without degrading chip lifetime, and thus save {approx}27% cooling power on average while providing the same processor reliability as a data center running at 288K.

97 MATHEMATICS AND COMPUTING↗

Fighting Fire with Fire: Superlattice Cooling of Silicon Hotspots to Reduce Global Cooling Requirements

The running costs of data centers are dominated by the need to dissipate heat generated by thousands of server machines. Higher temperatures are undesirable as they lead to premature silicon wear-out; in fact, mean time to failure has been shown to decrease exponentially with temperature (Black's law). Although other server components also generate heat, microprocessors still dominate in most server configurations and are also the most vulnerable to wearout as the feature sizes shrink. Even as processor complexity and technology scaling have increased the average energy density inside a processor to maximally tolerable levels, modern microprocessors make extensive use of hardware structures such as the load-store queue and other CAM-based units, and the peak temperatures on chip can be much worse than even the average temperature of the chip. In recent studies, it has been shown that hot-spots inside a processor can generate {approx} 800W/cm{sup 2} heat flux whereas the average heat flux is only 10-50W/cm{sup 2}, and due to this disparity in heat generation, the temperature in hot spots may be up to 30 C more than average chip temperature. The key problem processor hot-spots create is that in order to prevent some critical hardware structures from wearing out faster, the air conditioners in a data center have to be provisioned for worst case requirements. Worse yet, air conditioner efficiencies decrease exponentially as the desired ambient temperature decreases relative to the air outside. As a result, the global cooling costs in data centers, which nearly equals the IT equipment power consumption, are directly correlated with the maximum hot spot temperatures of processors, and there is a distinct requirement for a cooling technique to mitigate hot-spots selectively so that the global air conditioners can operate at higher, more efficient, temperatures. We observe that localized cooling via superlattice microrefrigeration presents exactly this opportunity whereby hot-spots can be cooled selectively and allow global coolers to operate at much more efficient temperatures. Recent advances in processor cooling technologies have demonstrated that thermoelectric coolers (TEC), which use a Peltier effect to form heat pumps, can be used to reduce the temperature of hot spots. By applying a thermoelectric cooler between the heat spreader and the processor die and applying current selectively at the hot spots, heat from the hot-spots can be spread much more efficiently. The ability to implement such thermoelectric coolers on a real silicon device has been demonstrated recently, albeit for small prototype chips. The key question then, that needs to be answered before such thermoelectric coolers can be integrated in commodity server processors, is 'What is the potential for superlattice microrefrigeration to reduce global cooling costs in data centers?'. In order to answer this question, we present a comprehensive analysis of the impact of thermoelectric coolers on global cooling costs. Our thermal analysis covers all aspects of cooling a server in a data center, and integrates on-chip dynamic and leakage power sources with a detailed heat diffusion model of a processor (that models the silicon to the thermoelectric cooler to the heat spreader and the heat sink) and finally the computer room air conditioner (CRAC) efficiency, as shown in Figure 1. In Section II, we present the components of the system model.

99 GENERAL AND MISCELLANEOUS↗