Developer finds parallel connections hurt speed test accuracy, not help it

A developer building a custom internet speed test discovered it was reporting 8.2 Mbps while Speedtest.net showed 28.65 Mbps on the same line. Initial suspicion pointed to server distance — the custom tool routed to Singapore at 120 ms versus a local Karachi server — but curl tests revealed the real culprit was parallel connection handling. On an already congested international link, opening more TCP connections triggered simultaneous congestion-control backoffs, collapsing total throughput below what a single connection achieved. The fix involved running two back-to-back download tests — one with two connections, one with eight — and reporting whichever performed better, which raised the measured speed to 15.1 Mbps. The developer also added a 600 ms drain period between test segments to prevent residual congestion from skewing results.