Architectural approach to the role of optics in monoprocessor and multiprocessor machines

Abstract

The relevance of introducing optical interconnects (OI's) in monoprocessors and multiprocessors is studied from an architectural point of view. We show that perhaps the major explanation for why optical technologies have nearly been unable to penetrate into computers is that OI's generally do not shorten the memory-access time, which is the most critical issue for today's stored-program machines. In monoprocessors the memory-access time is dominated by the electronic latency of the memory itself. Thus implementing OI's inside the memory hierarchy without changing the memory architecture cannot dramatically improve the global performance. In strongly coupled multiprocessors the node-bypass latency dominates. Therefore the higher the connectivity (possibly with optics), the shorter the path to another node, but the more expensive the network and the more complex the structure of electronic nodes. This relation leaves the choice of the best network open in terms of simplicity and latency reduction. The bottlenecks resulting from and the benefits of implementing OI's are discussed with respect to symmetric multiprocessors, rings, and distributed shared-memory supercomputers.

Reconfigurable and adaptive photonic networks for high-performance computing systems

“…Optical interconnects offer several well-known advantages for HPC systems such as higher spatial and temporal bandwidths, lower cross talk independent of data rates, higher interconnect densities, better signal integrity at high frequencies, lower signal attenuation, and lower power requirements at higher bit rates [2,3,[10][11][12][13][14], all of which could potentially achieve the much desired high bit rates data communication at a much reduced power level at the boardto-board distances of 0.1-1 m.…”

Predicting the performance of reconfigurable optical interconnects in distributed shared-memory systems

“…At the systems level, Collet [3], Huang [13] and Benner [4] each explore the potential of optical interconnections in several types of computing systems. All agree that optics can be used effectively in the class of tightly coupled, medium to large scale SMP systems.…”

“…11 which shows the complementary cumulative distribution of congestion over the links: for the baseline the line drops off slower meaning there are more links with higher congestion. 3 Further analysis of the impact of congestion is therefore necessary; we leave this for future work. Figure 12 shows the computation times required for both the full-system simulations and our prediction model, the former taking several hours while the latter can be completed in just a few minutes.…”

“…Figure 12 shows the computation times required for both the full-system simulations and our prediction model, the former taking several hours while the latter can be completed in just a few minutes. We did not include the cost of the initial simulations in the computation time for our method, since this 3 We define the congestion of a link as the aggregate time packets need to wait in a buffer immediately preceding this link before being transmitted over the link. It would be in a time × numbero f packets unit, but for the graph it was rescaled to an arbitrary range.…”

“…It has been shown that optical interconnection technologies can alleviate this bottleneck [3,4]. Mostly unhindered by crosstalk, attenuation, and increased capacitive bus load, these technologies will soon provide a cheaper, faster, and smaller alternative to electrical interconnections, on distances from a few centimeters upward.…”