Increasing the Effectiveness of Directory Caches by Avoiding the Tracking of Noncoherent Memory Blocks

Abstract

Abstract-A key aspect of the design of efficient multiprocessor systems is the cache coherence protocol. Although directory-based protocols constitute the most scalable approach, the limited size of the directory caches together with the growing size of systems may cause frequent evictions and, consequently, the invalidation of cached blocks, which jeopardizes system performance. Directory caches keep track of every memory block stored in processor caches in order to provide coherent access to the shared memory. However, a significant fraction of the cached memory blocks do not require coherence maintenance (even in parallel applications) because they are either accessed by just one processor or they are never modified. In this paper, we propose to deactivate the coherence protocol for those blocks that do not require coherence. This deactivation allows directory caches not to keep track of non-coherent blocks, which reduces directory cache occupancy and increases its effectiveness. Since the detection of non-coherent blocks is carried out by the operating system, our proposal only requires minor hardware modifications. Simulation results show that, thanks to our proposal, directory caches can avoid the tracking of about 66% of the blocks accessed by a wide range of applications, thereby improving the efficiency of directory caches. This contributes either to shorten the runtime of parallel applications by 15% while keeping directory cache size or to maintain performance while using directory caches 16 times smaller.

A Dual-Consistency Cache Coherence Protocol

“…While previous proposals classified memory accesses for guiding data placement [7], [13] or optimizing coherence protocols (e.g., reducing self-invalidation [2], reducing directory pressure [11], [21]), SPEL relies on the delineation of xDRF regions to minimize the coherence maintenance while providing SC guarantees, for any code. Next we review the most widely adopted approaches for memory access classification.…”

Design of Efficient TLB-based Data Classification Mechanisms in Chip Multiprocessors

“…Since data access patterns change throughout different phases of the application lifetime [39,78], adaptivity is fundamental to achieve a good classification accuracy. Read-only classification (i.e., detecting non-written regions of data) is also a far-reaching property, as merely accounting for shared-read-only blocks they can reach in some cases up to 48.7% of all accessed blocks [23]. Most data classification mechanisms can detect read-only data at the cost of raising the hardware complexity, increasing the network consumption or adding storage requirements.…”

“…For instance, Cuesta et al propose Coherence Deactivation, which identifies private [22] and readonly [23] (non-coherent) blocks, and avoids storing those blocks in the directory cache, since they do not require coherence maintenance. Therefore, directories exploit their limited storage capacity more efficiently and their access latency is reduced.…”

“…Thus, if a block is known to be present only in a single private cache (i.e., private), coherence maintenance may be bypassed. In this regard, Coherence Deactivation identifies private [22] and read-only [23] (non-coherent) blocks, and avoids storing those blocks in the directory cache. Therefore, directories exploit their limited storage capacity more efficiently as long as classification mechanism becomes more accurate, while the availability of the directory is improved for blocks that really need coherence (i.e., shared blocks).…”

Design of Efficient TLB-based Data Classification Mechanisms in Chip Multiprocessors

“…Non-adaptive approaches may miss-classify most accessed data as shared at some point in applications running for a long time, ultimately neglecting the advantages of the classification. Discern store memory operations allows determine read-only data, which represents a 82 percent (on average) of the memory blocks accessed [23]. Therefore, detecting read-only data significantly improves the potential of a classification mechanism.…”