Separate latency, throughput and loss by measurement

short · 50 min · Objective 6.2

Before you start — what this lab needs

Two Linux hosts on your own isolated lab network. Everything is equipment you own. iperf3 on both (iperf3 -s on host B), and tc from iproute2 on host A, which shapes its own outgoing traffic:

tc qdisc add dev IFACE root netem delay 100ms       # step 2
tc qdisc change dev IFACE root netem loss 1%        # step 3 (delay gone, loss on)
tc qdisc replace dev IFACE root tbf rate 50mbit burst 32kbit latency 400ms   # step 4
tc qdisc del dev IFACE root                         # back to normal

Run the iperf3 CLIENT on host A. tc only shapes what A sends, and the client is the sender: shape the receiver instead and the loss lands on ACKs, which TCP barely notices.

Task

Introduce each of the three network impairments independently and measure how they differ, then show the relationship that confuses people: high latency limits throughput even when bandwidth is plentiful.

Steps

Measure the baseline from host A: round-trip time and loss with ping -c 200 -i 0.2 B, single-stream throughput with iperf3 -c B -t 30, and eight streams with iperf3 -c B -t 30 -P 8. Record in lab/net/impair.csv, one row per condition with columns condition,rtt_ms,single_mbps,multi_mbps,loss_pct.
Add 100 ms of latency only -- no loss, no bandwidth limit. Re-measure all four and record. Note what happened to single-stream throughput despite the link being unchanged. A current Linux stack auto-tunes its window up to several megabytes, which covers 100 ms at a few hundred Mbit/s; if the drop is small on your link, add a row with iperf3 -w 256K, a window an untuned or older host would use.
Remove the latency and add 1% packet loss only. Re-measure and record the effect on throughput.
Remove the loss and impose a hard bandwidth cap. Re-measure and record how this looks different from the first two.
With the 100 ms latency restored, compare single-stream against eight-stream throughput and write in lab/net/impair.md why parallel streams helped and why a bigger link would not have.

Verify

awk -F, 'NR>1 && NF>=4 {n++} END {print n" measurement set(s)"}' lab/net/impair.csv
grep -Eci 'window' lab/net/impair.md
grep -Eci 'parallel|multi.stream' lab/net/impair.md
grep -Eci 'loss' lab/net/impair.md

Four measurement sets, and the notes must explain the window/RTT relationship. A small amount of loss should hurt throughput out of proportion to its size -- if it did not, the transfer was too short to trigger congestion control.

This is an independent study companion for CompTIA Cloud+ CV0-004 and is not produced by or endorsed by CompTIA.