TY - JOUR
T1 - HFOS L
T2 - hyper scale fast optical switch-based data center network with L-level sub-network
AU - Khani, Elham
AU - Hessabi, Shaahin
AU - Koohi, Somayyeh
AU - Yan, Fulong
AU - Calabretta, Nicola
PY - 2022/7
Y1 - 2022/7
N2 - The ever-expanding growth of internet traffic enforces deployment of massive Data Center Networks (DCNs) supporting high performance communications. Optical switching is being studied as a promising approach to fulfill the surging requirements of large scale data centers. The tree-based optical topology limits the scalability of the interconnected network due to the limitations in the port count of optical switches and the lack of optical buffers. Alternatively, buffer-less Fast Optical Switch (FOS) was proposed to realize the nanosecond switching of optical DCNs. Although FOSs provide nanosecond optical switching, they still suffer from port count limitations to scale the DCN. To address the issue of scaling DCNs to more than two million servers, we propose the hyper scale FOS-based L-level DCNs (HFOSL) which is capable of building large networks with small radix switches. The numerical analysis shows L of 4 is the optimal level for HFOSL to obtain the lowest cost and power consumption. Specifically, under a network size of 160,000 servers, HFOS4 saves 36.2% in cost compared with the 2-level FOS-based DCN, while achieves 60% improvement for cost and 26.7% improvement for power consumption compared with Fat tree. Moreover, a wide range of simulations and analyses demonstrate that HFOS4 outperforms state-of-art FOS-based DCNs by up to 40% end-to-end latency under DCN size of 81920 servers.
AB - The ever-expanding growth of internet traffic enforces deployment of massive Data Center Networks (DCNs) supporting high performance communications. Optical switching is being studied as a promising approach to fulfill the surging requirements of large scale data centers. The tree-based optical topology limits the scalability of the interconnected network due to the limitations in the port count of optical switches and the lack of optical buffers. Alternatively, buffer-less Fast Optical Switch (FOS) was proposed to realize the nanosecond switching of optical DCNs. Although FOSs provide nanosecond optical switching, they still suffer from port count limitations to scale the DCN. To address the issue of scaling DCNs to more than two million servers, we propose the hyper scale FOS-based L-level DCNs (HFOSL) which is capable of building large networks with small radix switches. The numerical analysis shows L of 4 is the optimal level for HFOSL to obtain the lowest cost and power consumption. Specifically, under a network size of 160,000 servers, HFOS4 saves 36.2% in cost compared with the 2-level FOS-based DCN, while achieves 60% improvement for cost and 26.7% improvement for power consumption compared with Fat tree. Moreover, a wide range of simulations and analyses demonstrate that HFOS4 outperforms state-of-art FOS-based DCNs by up to 40% end-to-end latency under DCN size of 81920 servers.
KW - Fast optical switch
KW - Low latency interconnection
KW - Optical data center network
KW - Scalable data center network
UR - http://www.scopus.com/inward/record.url?scp=85130263189&partnerID=8YFLogxK
U2 - 10.1007/s11235-022-00905-2
DO - 10.1007/s11235-022-00905-2
M3 - Article
AN - SCOPUS:85130263189
SN - 1018-4864
VL - 80
SP - 397
EP - 411
JO - Telecommunication Systems
JF - Telecommunication Systems
IS - 3
ER -