Network Performance

• Latency
• Throughput
• Internet Performance
• High Speed Networks

Latency

• Latency: from point A to point B
• Round-Trip Time: from point A to point B and back to point A
• RTT is much easier to measure, though with GPS even latency is becoming possible
• latency on one link is t = t_s + t_d + t_q:
  • the time to put data on the wire: given bandwidth b and packet size s, t_s = s/b.
  • the time for the data to travel at the speed of light the distance d to the destination: t_d = d/c_medium
  • the time the packet is queued before being sent: t_q
• for a given size packet, the only variable is t_q

Multi-Hop Latency

• Measuring Link Bandwidths Using a Deterministic Model of Packet Delay (Lai and Baker, SigComm 2000):
  • l is the link number, s the packet size, b_l the bandwidth of link l
  • in the absence of queueing delays (q_d = 0), t_l = t₀ + \sum_i=0^l-1({s / {b_i}} + d_i)
• with wormhole routing, replace s/b_i with min(0, s/b_i - s/b_i-1)
• if t_q = 0, and if for some x the bandwidth b_x << b_l for all other links l, then t =~ s/b_l
• if t_q = 0, and all n links have the same bandwidth b, then t =~ ns/b_l

Throughput

• Send throughput: for "send only", only tells you about the speed of the sending machine
• Receive throughput: how many bits per second we received, given that the sender was sending as fast as possible
• "Goodput": received correctly, by application
• have to define terms carefully: do you count retransmissions? packet losses? erroneous packets?
• due to congestion management, TCP throughput may be much less than send/receive throughput tested using UDP
• especially true in the case of lost packets or packets with errors

Window-based Flow Control

• if the bandwidth-delay product is bigger than the window, the protocol and not the network will be the bottleneck
• high-bandwidth, (relatively) high-delay networks
• example: 1Gb/s cross-country (100ms) link
• TCP window scaling
• reusing TCP sequence numbers: TCP timestamps (RFC 1323)
• Selective Retransmission without going back to window size of 1 MSS (TCP Fast Retransmit)

Internet Performance

• some connections are great
• some destinations are permanently unreachable from others
• some destinations (e.g. volcano from UH) have a high error rate (typically 4% of packets lost)
• congestion is unpredictable
• throughput is unpredictable
• web page caching (e.g. project 3, Akamai) helps "perceived" throughput

High Speed Networks

• Gigabit Ethernet: 1Gb/s
• HSSI (High Speed Serial Interface): only 0.052 Gb/s
• HiPPI (High Performance Parallel Interface): 0.8 - 1.6 Gb/s
• Satellite: high bandwidth, high delay
• Microwave
• SONET: OC-12 (0.622 Gb/s)
• ATM: OC-12 (0.622 Gb/s)

High Speed Home Networks

• Cable Modem: 10Mb/s Ethernet on a reserved frequency band ("channel")
• 56Kb/s modems: digital from the ISP to the central office, analog to the home
• (A)DSL: using better (smarter) modems to achieve speeds higher than 128Kb/s (forward error correction)

Overall Performance Issues

• time is what people really care about:
  • RTT for real-time traffic
  • response time
  • overall time to complete a task
• some combination of a * RTT + b * bandwidth
• the math matters!