A New Memory Module for Memory Intensive Applications

Abstract

Some applications with gather / scatter operations are difficult to accelerate. These operations cause inefficient cache use in each processor and fine grain global communications in parallel systems. There are several applications with such characteristics particularly in electrical engineering. For examples, circuit simulation and power flow simulation with LU decomposition for random sparse matrix has such characteristics. This paper presents how to make inexpensive personal supercomputers to solve these problems. In order to get the merit of commercial-off-the-shelf (COTS) continuously after the death of vector supercomputer vendors, it is designed without any modification on CPU, bridge chips on motherboard and memory chips. Only plugging a new memory module with vector load / store function and communication functions make an inexpensive home-use personal computer into a node similar to Earth simulator's one. Applications with unit striding or indexed accesses are going to be accelerated. How to accelerate NAS CG is shown as an example.

An enhancer of memory and network for applications with large-capacity data and non-continuous data accessing

“…In order to improve performance for applications with discontinuous access in the conventional Intel-based PC, we proposed vector scatter and gather access functions on a memory module such as DIMMnet-2 [19,20]. This concept can be adopted for an AMD-based PC with an attached accelerator on a Hypertransport-based CPU socket such as Reconfigurable Processor Unit (RPU110) [9] made by DRC Computer Corporation.…”

“…Tanabe named this buffer "window memory" in the previous proposed memory module, DIMMnet-1 [18]. Figure 2 shows the window memory with vector access functions on DIMMnet-2 [19,20].…”

“…Our solution for these problems of COTS comprises vector load/store functions and capacity extensibility on a memory module such as DIMMnet-2 [19,20] or an attached accelerator on a CPU socket such as the DRC Reconfigurable Processor Unit (RPU110) [9]. Since, unlike Impulse, DIMMnet-2 does not suffer from the size limitation of CPU address space and no remanufacturing of a bridge chip is necessary, it can be utilized with a conventional COTS CPU, motherboard, and memory chip.…”

“…The basic concept of a memory module with vector load/store functions was proposed by Tanabe [19,20]. In Tanabe's work, the bandwidth effect was statically analyzed [19] and CG benchmark in small class (class B) for idealized hardware was evaluated [20] preliminarily.…”

“…In Tanabe's work, the bandwidth effect was statically analyzed [19] and CG benchmark in small class (class B) for idealized hardware was evaluated [20] preliminarily. The impulse memory controller was evaluated with CG in very small class (class A) by Carter [5].…”