HyStart testing between Princeton and NCSU

From Wiki

Setup

• We tested HyStart in the path between Princeton and NCSU. An estimated bandwidth for this path is 100Mbps and the RTT is 25ms. We tested CUBIC with two clock resolutions (HZ=100 and HZ=1000). The buffer auto-tuning was enabled. We applied three patches in addition to the recent patches (PATCH 0/6 TCP CUBIC and Hystart) from Stephen Hemminger.

• Tested protocols

 CUBIC with SlowStart (by disabling HyStart)
 CUBIC with HyStart (without the patches) 
 CUBIC with HyStart (with the patches)

• Installation

 Download the latest git from Linus.
 Apply six patches from Stephen Hemminger ([1])
 Apply these three patches[2][3][4]

• Credits

 Patches enabling high-resolution ACK time (if HZ < 1000) from Stephen Hemminger
 Problem reporting from Lucas Nussbaum
 Testing CUBIC from Bill Fink
 Discussion with Injong Rhee

Results

• When the clock resolution is coarse (e.g., HZ < 1000), the accuracy of HyStart is affected and it may exit slow start prematurely.
• Even though Hystart found a safe exit point, it may be still slow to reach to maximum available bandwidth in long BDP paths due to initial slow growth of CUBIC.
• The above three patches solve this problem.

Closer look at the first 20 secs

Clock / Protocols	HZ = 100	HZ = 1000
Slowstart
HyStart (before fix)
HyStart (fixed)