Components of a "Cayley" data center.[/caption] A team of researchers from Microsoft and Cornell University has concluded that, in some cases, a totally wireless data center makes logistical sense. In a new paper, a team of researchers from Cornell and Microsoft concluded that a data-center operator could replace hundreds of feet of cable with 60-GHz wireless connections—assuming that the servers themselves are redesigned in cylindrical racks, shaped like prisms, with blade servers addressing both intra- and inter-rack connections. The so-called “Cayley” data centers, so named because of the network connectivity subgraphs are modeled using Cayley graphs, could be cheaper than traditional wired data centers if the cost of a 60-GHz transceiver drops under $90 apiece, and would likely consume about one-tenth to one-twelfth the power of a wired data center. There’s just one problem, however: Cayley datacenters are expected to show significantly better latency on average than conventional data centers and so-called “fat tree” networks, except under peak load. “Conventional datacenters, based on wired networks, entail high wiring costs, suffer from performance bottlenecks, and have low resilience to network failures,” the paper’s authors, Ji-Yong Shin, Emin Gün Sirer, and Hakim Weatherspoon of Cornell, wrote. The co-author was Darko Kirovski, of Microsoft Research. “Our exploration of the resulting design space shows that wireless datacenters built with this methodology can potentially attain higher aggregate bandwidth, lower latency, and substantially higher fault tolerance than a conventional wired datacenter while improving ease of construction and maintenance,” they added. Although many 60-GHz technologies are under consideration (IEE 802.15.3c and 802.11ad, WiGig, and others), the authors picked a Georgia Tech design with bandwidth of between 4-15Gbps and and effective range of less than or equal to 10 meters. Beam-steering wasn’t used because of the latencies involved in reinstating a dropped connection, although both time and frequency multiplexing were. (Because the team couldn’t actually build the design, they chose Terabeam/HXI 60-GHz transceivers for a conservative estimate.) During their testing, the authors arranged the servers in a circular pattern so that inter- and intra-rack communication channels could be established and form a densely connected mesh. As part of that, they replaced the traditional NIC with a “Y-switch” that connects a server’s system bus with two transceivers positioned at opposite ends of the server box. “This topology leads to full disappearance of the classic network switching fabric (e.g., no top-of-rackswitches, access routers, copper and optical interconnects) and has far-reaching ramifications on performance,” they wrote. In all, the researchers found, the Cayley datacenters can maintain connectivity to over 99 percent of live nodes until up to 55 percent of total nodes fail. The authors acknowledged that considerable work was needed to take the project further, if only to answer questions such as whether an additional wireless network could address local congestion and MAC issues, if an alternate wired network would benefit performance, or if it would be beneficial to parallelize the system into a substantially larger number of low-power low-cost less-powerful processors and support hardware. Image: "On the Feasibility of Completely Wireless Datacenters" Is a Wireless Data Center Possible?
[caption id="attachment_5201" align="aligncenter" width="447"] Components of a "Cayley" data center.[/caption] A team of researchers from Microsoft and Cornell University has concluded that, in some cases, a totally wireless data center makes logistical sense. In a new paper, a team of researchers from Cornell and Microsoft concluded that a data-center operator could replace hundreds of feet of cable with 60-GHz wireless connections—assuming that the servers themselves are redesigned in cylindrical racks, shaped like prisms, with blade servers addressing both intra- and inter-rack connections. The so-called “Cayley” data centers, so named because of the network connectivity subgraphs are modeled using Cayley graphs, could be cheaper than traditional wired data centers if the cost of a 60-GHz transceiver drops under $90 apiece, and would likely consume about one-tenth to one-twelfth the power of a wired data center. There’s just one problem, however: Cayley datacenters are expected to show significantly better latency on average than conventional data centers and so-called “fat tree” networks, except under peak load. “Conventional datacenters, based on wired networks, entail high wiring costs, suffer from performance bottlenecks, and have low resilience to network failures,” the paper’s authors, Ji-Yong Shin, Emin Gün Sirer, and Hakim Weatherspoon of Cornell, wrote. The co-author was Darko Kirovski, of Microsoft Research. “Our exploration of the resulting design space shows that wireless datacenters built with this methodology can potentially attain higher aggregate bandwidth, lower latency, and substantially higher fault tolerance than a conventional wired datacenter while improving ease of construction and maintenance,” they added. Although many 60-GHz technologies are under consideration (IEE 802.15.3c and 802.11ad, WiGig, and others), the authors picked a Georgia Tech design with bandwidth of between 4-15Gbps and and effective range of less than or equal to 10 meters. Beam-steering wasn’t used because of the latencies involved in reinstating a dropped connection, although both time and frequency multiplexing were. (Because the team couldn’t actually build the design, they chose Terabeam/HXI 60-GHz transceivers for a conservative estimate.) During their testing, the authors arranged the servers in a circular pattern so that inter- and intra-rack communication channels could be established and form a densely connected mesh. As part of that, they replaced the traditional NIC with a “Y-switch” that connects a server’s system bus with two transceivers positioned at opposite ends of the server box. “This topology leads to full disappearance of the classic network switching fabric (e.g., no top-of-rackswitches, access routers, copper and optical interconnects) and has far-reaching ramifications on performance,” they wrote. In all, the researchers found, the Cayley datacenters can maintain connectivity to over 99 percent of live nodes until up to 55 percent of total nodes fail. The authors acknowledged that considerable work was needed to take the project further, if only to answer questions such as whether an additional wireless network could address local congestion and MAC issues, if an alternate wired network would benefit performance, or if it would be beneficial to parallelize the system into a substantially larger number of low-power low-cost less-powerful processors and support hardware. Image: "On the Feasibility of Completely Wireless Datacenters"
Components of a "Cayley" data center.[/caption] A team of researchers from Microsoft and Cornell University has concluded that, in some cases, a totally wireless data center makes logistical sense. In a new paper, a team of researchers from Cornell and Microsoft concluded that a data-center operator could replace hundreds of feet of cable with 60-GHz wireless connections—assuming that the servers themselves are redesigned in cylindrical racks, shaped like prisms, with blade servers addressing both intra- and inter-rack connections. The so-called “Cayley” data centers, so named because of the network connectivity subgraphs are modeled using Cayley graphs, could be cheaper than traditional wired data centers if the cost of a 60-GHz transceiver drops under $90 apiece, and would likely consume about one-tenth to one-twelfth the power of a wired data center. There’s just one problem, however: Cayley datacenters are expected to show significantly better latency on average than conventional data centers and so-called “fat tree” networks, except under peak load. “Conventional datacenters, based on wired networks, entail high wiring costs, suffer from performance bottlenecks, and have low resilience to network failures,” the paper’s authors, Ji-Yong Shin, Emin Gün Sirer, and Hakim Weatherspoon of Cornell, wrote. The co-author was Darko Kirovski, of Microsoft Research. “Our exploration of the resulting design space shows that wireless datacenters built with this methodology can potentially attain higher aggregate bandwidth, lower latency, and substantially higher fault tolerance than a conventional wired datacenter while improving ease of construction and maintenance,” they added. Although many 60-GHz technologies are under consideration (IEE 802.15.3c and 802.11ad, WiGig, and others), the authors picked a Georgia Tech design with bandwidth of between 4-15Gbps and and effective range of less than or equal to 10 meters. Beam-steering wasn’t used because of the latencies involved in reinstating a dropped connection, although both time and frequency multiplexing were. (Because the team couldn’t actually build the design, they chose Terabeam/HXI 60-GHz transceivers for a conservative estimate.) During their testing, the authors arranged the servers in a circular pattern so that inter- and intra-rack communication channels could be established and form a densely connected mesh. As part of that, they replaced the traditional NIC with a “Y-switch” that connects a server’s system bus with two transceivers positioned at opposite ends of the server box. “This topology leads to full disappearance of the classic network switching fabric (e.g., no top-of-rackswitches, access routers, copper and optical interconnects) and has far-reaching ramifications on performance,” they wrote. In all, the researchers found, the Cayley datacenters can maintain connectivity to over 99 percent of live nodes until up to 55 percent of total nodes fail. The authors acknowledged that considerable work was needed to take the project further, if only to answer questions such as whether an additional wireless network could address local congestion and MAC issues, if an alternate wired network would benefit performance, or if it would be beneficial to parallelize the system into a substantially larger number of low-power low-cost less-powerful processors and support hardware. Image: "On the Feasibility of Completely Wireless Datacenters" 