Cloud Computing requires a holistic look at network performance. Our customers and Cloud network architects don’t design to one singular metric for cloud containerization. Its a balanced cloud for predictable/low latency (in nano secs to few microseconds), non blocking 1 or 10GE throughput, guaranteed packet delivery/buffers and appropriate uplinks. Not to mention the need for resilience and open extensibility.
Jayshree Ullal
CEO, Arista Networks
Using FocalPoint in the Data Center
As the size and density of data centers increase, the cost, area, power and support of multiple interconnect fabric technologies cannot be tolerated. Because of this, the IEEE is developing several new standards that will enable Ethernet as the single unified fabric for data, storage and HPC traffic. The industry has coined these new initiatives Converged Enhanced Ethernet (CEE) or Data center Bridging (DCB). FocalPoint has been developed with multiple features to support CEE-DCB applications. The advanced features available in the FocalPoint switch family also help data center designers cope with the problems of scalability, space, power and cost as they evolve to meet increasing demands for these applications.
Clustering
FocalPoint switches provide this scalability with up to 24 ports of 10 Gigabit Ethernet connectivity – the most highly integrated 10 Gigabit Ethernet switch available. If even higher bandwidth connections are needed, Ethernet’s native Link Aggregation Protocol provides the ability to aggregate multiple 10 Gigabit ports into a single, higher bandwidth pipe. With FocalPoint, link aggregation can be done even across multiple switches in a multi-stage topology such as a fat tree or stack as shown in the figure below. By using standard Ethernet switching throughout the fabric, scalability is achieved that is more cost, space and power effective than the standard “Big Iron” approach while maintaining less that 1uS latency through three switch stages.
Network Convergence and Storage
Network convergence can reduce overall system cost by combining networking, storage and HPC traffic onto one network. As the central nervous system of the data center, the switch/router network must effectively deal with all three types of traffic at once, and be able to distinguish between them, allocating its resources intelligently among them. FocalPoint provides multiple priority-based classes per port, enabling the system administrator to finely tune the transmission characteristics of each type of traffic. Advanced scheduling, shaping and policing algorithms also allow the administrator to monitor and enforce bandwidth allocation among different classes of users.
Storage protocols such as FCoE require deterministic, bounded latency to ensure that time-out values are not exceeded. Traditional Ethernet switches drop packets during periods of high congestion, requiring retransmission at a higher layer. This added latency cannot be tolerated in storage applications. Standard IEEE pause frames will pause all traffic during periods of congestion. So with large enough buffers, traffic will not be dropped, but unacceptable latency times may still be seen. Because of this, the IEEE is standardizing Priority Flow Control (PFC), which can differentiate traffic classes, providing bounded latency for storage traffic. FocalPoint can classify and assign storage traffic to a special traffic class and memory partition, which combined with PFC provides lossless operation.
In any switch fabric, traffic from multiple ingress ports can compete for the limited bandwidth in a single egress port. For example, bursts of data traffic could reduce the bandwidth available to storage traffic, causing congestion and increased latency for the storage traffic. To solve this problem, the IEEE is developing Enhanced Transmission Selection (ETS) which provides advanced traffic scheduling including features such as guarantee minimum bandwidth for certain traffic classes like storage or HPC. FocalPoint supports a superset of the features required for ETS which can result in bounded latency when required.
In multi-stage fabrics, congestion hot spots can occur which can cause congestion spreading. To combat this, a new IEEE work group has been formed to develop the Quantized Congestion Notification (QCN) standard. FocalPoint was developed with several features to support the QCN standard, which have been tested in the Monaco reference platform. By using QCN, fabric cross-sectional bandwidth can be optimized, minimizing the need for additional switches while taking advantage of the scalability, low-latency available with FocalPoint.
