Level 5 Networks: EtherFabric

EtherFabric performance measurements demonstrate the dramatic innovation contained in the architecture. Using industry-standard performance measures, EtherFabric is clearly a significant leap forward from standard Ethernet. It is also competitive with proprietary high-performance interconnects that were once the only way to get good application performance.

CPU Efficiency – EtherFabric delivers more computing power to the application

CPU Efficiency: CPU efficiency provides 3.5 times the number of CPU cycles

CPU efficiency is the key performance measure of any interconnect. Efficient handling of network traffic frees CPU cycles to accelerate application performance. In networked servers, a large percentage of CPU cycles can be used for network communication, thereby limiting the overall performance of the application. CPU efficiency must always be measured up to and including the application actually receiving the data.

EtherFabric at least doubles the CPU cycles available to the applications on any server it is plugged into. This is shown above in measurements of the NetPerf benchmark.

Latency – EtherFabric dramatically reduces networked server latency

Latency: Server latency is reduced by 5 times to fewer than 10 µsecs

In many instances, latency is a very important determinant of application performance in data centers and compute clusters. It is the fundamental delay incurred in data transitioned from one element to another in the data center, and eventually to the user. This delay is often compounded in the data center environment since responding to a user may entail many element transitions, that is, from the load balancer – to the web tier server – to the application tier server – to the database tier server – to the storage network, then back to the database tier server – back to the application tier server – back to the web tier server – and back to the use, each one of these connections passing through an appropriate Ethernet switch.

High latency has been a problem with conventional Gigabit Ethernet for a long time. Typically Ethernet latency is in the range of 50-100 microseconds. In contrast, EtherFabric offers sub 10 microsecond latency, which provides a dramatic improvement, especially in a data center or compute cluster environment as multiple equipment intercommunications occur.

Bandwidth – EtherFabric doubles the amount of useable bandwidth
to the application

Bandwidth: Applications have 91% more network bandwidth

Bandwidth is another key performance determinant of applications running in an interconnected data center. The effect on performance of low bandwidth is apparent when the data being passed between equipment is comprised of large messages. With small messages, which are customary for such activities as web inquiries, transaction processing and the like, bandwidth is not limiting and latency more strongly influences performance. However, in large message scenarios such as storage, streaming media or block transfer communication, a lack of bandwidth can be damaging to overall performance. And because multiple transfers are needed to handle one message, bandwidth limits are reached and latency is increased.

In the figure, it can be seen that EtherFabric provides much greater bandwidth at every message size. Additionally, the port striping capability enables EtherFabric to deliver beyond 1 Gb; in fact, delivering close to the theoretical wire throughput of both ports - 1.9Gb. And at this throughput, ample CPU cycles area vailable to the application to enable effective use of the bandwidth. In contrast, conventional Ethernet tops out at 1.3Gb when bonding is attempted, and 100% of the CPU is utilized, making the bandwidth useless in any event.