Exascale systems are in need for alternative interconnection technologies. Electrical interconnects are not likely to scale well to a large number of computing nodes in terms of energy efficiency and latency. Silicon photonic networks stand as the main alternative to solve this problem, but arewe there yet? In this paper, we exhibit some challenges to be solved for this technology to become a viable solution. Signal loss sources play a critical role in photonic network designs as they restrict the ability to perform data transmission in an effective way.
KEYWORDS
networks, optical interconnects, photonics
INTRODUCTIONProgress in scientific fields, including clime, aerospace, biotechnology, and energy, depends largely on the ability to perform costly and complex simulations. Supercomputers are the only viable option to support such computations, and intense research is focusing on increasing the computing capability of these systems. In fact, the main objective is the design of exascale systems, for which it becomes necessary to greatly increase the number of compute nodes. This raises numerous challenges that must be solved to obtain an efficient system in terms of cost, energy consumption, and performance. Some of these challenges are: scalable system software, resilience and correctness, programming systems, energy efficiency, or interconnect technology.A significant growth in parallelism implies that the system performance is greatly determined by the communication generated when running the parallel applications, even more than the arithmetic operations. Note that data transfers exist at several levels: between compute and storage nodes, between compute nodes, and between the multiple computing resources of each multicore node. Therefore, data movement is a critical barrier toward realizing the exascale systems, and thus, the interconnection network is a key component of exascale systems.Moving to exascale systems seems that it will not be possible with a traditional incremental strategy, and significant qualitative changes will be necessary. In the case of the interconnect systems, electrical interconnects are not likely to scale well to a large number or computing nodes in terms of energy efficiency and latency 1 ; therefore, different technologies to the traditional ones could contribute to achieve this objective. Photonics is perhaps the most disruptive technology 2-5 due to its capabilities to transmit and receive high bandwidth signals with higher power efficiency and immunity to degradation.Significant progress has been made over the past decade in optical device integration. 6-8 However, photonic devices are different in how they function, and exploiting all its advantages would require a significant change in how on-and off-chip interconnects are designed. Nevertheless, this paradigm has some challenges to be addressed, for example, optical signal buffering is greatly limited by the write speed of data. 9 Schemes keeping the writings to its minimum would be needed. Therefore, new proposals to solve t...
