Packet throughput is the number of bits per second (bps) or packets per second (pps) sent and received between source and destination. There are three primary types of throughput that affect network performance. They include network, device and link throughput. Throughput is often expressed as a 5 minute average and fluctuates with network conditions. Aggregate throughput is affected by all the network devices between desktop and server. Typical components that affect network throughput include device capacity, scalability and network interface speed. Link throughput is the amount of data forwarded in bps or pps during a fixed period across a campus or WAN link. The links could include switch to switch, switch to router, switch to server and WAN circuits. The link throughput is determined by the interface and link speed for serializing bits and forwarding them across an uplink. Link bandwidth utilization is a metric that is calculated using throughput and link speed.
Device throughput is expressed as aggregate switching capacity (bps). The devices are often rated with backplane fabric capacity and inter-module switching capacity. This is often more important than a performance perspective to switches that aggregate traffic at distribution and core switches.
Network throughput is the aggregate of both link and device throughput. Network device throughput is often rated based on the worst case minimum 64 byte packet size. This results in an advertised throughput rating with the minimum throughput performance. The larger packet size typical of most network traffic will have a higher bps throughput rating. Packet size as a result has a significant effect on throughput. Most applications today use a variety of packet sizes with the average size much larger than 64 bytes.
TCP receive window (RWIN) size has the most significant effect on link and network throughput. The effect of TCP window size on WAN through is a good example. Packet throughput across a WAN circuit for instance is affected by the TCP window size and round trip time (RTT) latency. The TCP window size changes with fluctuating with network conditions. The default initial congestion TCP window size is 3 x MSS (4380 bytes). That is based on the RFC however that can vary among vendor TCP stack implementations.
The TCP receive window is increased to the traditional 65,535 byte maximum or until network congestion occurs. The amount of RTT latency is affected by fixed and variable network delays. The latency and throughput across the campus switched network at the data center is near Ethernet practical limits with 2 msec RTT latency. For instance using the maximum TCP RWIN of 65,535 bytes divided by network latency of 2 msec is approximately 262 Mbps. That is for a switching infrastructure that is Gigabit speed (1000 Mbps) The TCP maximum window size is the limiting factor for throughput. The TCP Window Scaling feature increases maximum TCP window size from 65,535 bytes to 1 GB (1,000,000,000 bytes) for high speed links. Note that Window Scaling is designed for high speed links with high latency for best results.
This is an example of the campus infrastructure throughput using the maximum TCP window size of 65,535 bytes and RTT latency of 2 msec.
campus throughput = TCP window size (bytes) x 8 bits / byte / RTT (sec)
= 65,535 bytes x 8 bits / byte / 0.040 sec
= 262 Mbps
= 26% utilization of a Gigabit link (1000 Mbps)
WAN Link Throughput
This is an example of WAN link throughput calculation using an average TCP window size of bytes and RTT latency of 150 msec. The primary cause of decreased throughput here is the RTT latency. The congestion on a slower T1 link will cause the smaller TCP window and that affects throughput.
WAN link throughput = TCP window size (bytes) / RTT (sec)
= 11680 bytes x 8 bits / byte / 0.150 sec
= 622,933 bps
= 40% utilization of a T1 circuit (1,544,000 bps)
Switch Device Throughput
This is an example of maximum forwarding rate calculation for a Cisco switch with 48 Gigabit ports. Cisco uses an 84 byte packet (frame) size for worst case performance calculations.
1,000,000,000 (GE) bps / 8 bits / byte = 125 Mbps
125 Mbps / 84 byte packet = 1.5 Mpps
forwarding rate = 1.5 Mpps x 48 GE ports = 72 Mpps
Copyright © 2013 Shaun L. Hummel All Rights Reserved