Die-Stacked DRAM: Memory, Cache, or MemCache?

Abstract

Die-stacked DRAM is a promising solution for satisfying the ever-increasing memory bandwidth requirements of multi-core processors. Manufacturing technology has enabled stacking several gigabytes of DRAM modules on the active die, thereby providing orders of magnitude higher bandwidth as compared to the conventional DIMM-based DDR memories. Nevertheless, die-stacked DRAM, due to its limited capacity, cannot accommodate entire datasets of modern big-data applications. Therefore, prior proposals use it either as a sizable memory-side cache or as a part of the software-visible main memory. Cache designs can adapt themselves to the dynamic variations of applications but suffer from the tag storage/latency/bandwidth overhead. On the other hand, memory designs eliminate the need for tags, and hence, provide efficient access to data, but are unable to capture the dynamic behaviors of applications due to their static nature.In this work, we make a case for using the die-stacked DRAM partly as main memory and partly as a cache. We observe that in modern big-data applications there are many hot pages with a large number of accesses. Based on this observation, we propose to use a portion of the die-stacked DRAM as main memory to host hot pages, enabling serving a significant number of the accesses from the high-bandwidth DRAM without the overhead of tag-checking, and manage the rest of the DRAM as a cache, for capturing the dynamic behavior of applications. In this proposal, a software procedure pre-processes the application and determines hot pages, then asks the OS to map them to the memory portion of the die-stacked DRAM. The cache portion of the die-stacked DRAM is managed by hardware, caching data allocated in the off-chip memory. Through detailed evaluations of various big-data applications, we show that our proposal improves system performance by 28% and up to 139% over the previous best-performing proposal.

MANA: Microarchitecting an Instruction Prefetcher

“…To reduce the storage cost, MANA exploits this observation that there are a small number of distinct high-order-bits patterns. This phenomenon is because the code base of a program has a high spatial locality and is much smaller than the size of the physical memory [9,15]. Consequently, instead of recording the complete trigger address that is the largest field of spatial regions, MANA uses the pointers to the observed high-order-bits patterns that need a considerably fewer number of bits.…”

A Survey on Recent Hardware Data Prefetching Approaches with An Emphasis on Servers

“…Progress in technology fabrication accompanied by circuit-level and microarchitectural advancements have brought about signi cant enhancements in the processors' performance over the past decades. Meanwhile, the performance of memory systems has not held speed with that of the processors, forming a large gap between the performance of processors and memory systems [29][30][31][32][33][34][35][36][37][38][39][40][41].…”

Building Chinese interdisciplinary research centers : the case of Tsinghua University

“…Figure 2 also highlights the changes in Trimma that adopt novel remap table and remap cache designs, which are merely in the metadata lookup and update phases (❶❹) without affecting the rest data access/eviction. Hence innovations including replacement policies [4], [27], [54], [64], [67], selective migration [16], [17], [82], and cache-flat dual modes [7], [10], [43] can be orthogonally integrated with Trimma.…”