HTTP / 2 Performance in Cellular Networks

A webpage with 376 objects of size 1 KB each. A webpage with 10 objects of size 435 KB each. A webpage with 136 objects of size 1 KB to 620 KB each. ● Page load times (PLTs) over h2 are lower than PLTs over h1, because in the case of h1 with 6 TCP connections the server can send at most 6 objects in parallel. ● Whereas in the case of h2 with many streams multiplexed onto single connection, the server sends a large number of objects in parallel. ● As the network condition becomes Very Poor, the PLTs increase for both h1 and h2, but more so for h2. For h1, the impact of packet loss on any one of the 6 connections only affects that connection. ● In the case of h2, since all object downloads are multiplexed over a single TCP connection, packet loss affects all ongoing object downloads ● About one-third of the total connections in cellular networks experience packet loss, as interpreted by TCP stack. ● Losses are often clustered, such that when a loss event occurs, many consecutive TCP segments are retransmitted by server. ● TCP connections experience retransmissions multiple times during their lifetimes. Subsequent retransmissions often appear within a one second interval. ● When a retransmission event occurs, many connections experience about 5-15% packet loss. There is currently no known best practice as to how TCP and HTTP/2 should be tuned such that the impact of loss on single h2 connection is minimized. We propose to side-step from the recommendation of using single HTTP/2 connection and investigate the mobile Web performance with multiple HTTP/2 connections. ● h1 outperforms h2 across all network conditions, especially in Poor conditions where loss occurs frequently. ● For h2 with one connection, the initial congestion window (ICW) size of the server during the TCP slow start is only one-sixth of the cumulative window sizes of h1 with 6 TCP connections. ● Thus, the server sends six times less data over h2 during the TCP slow start phase. ● As the network conditions worsen, the congestion window of the single h2 connection does not grow. ● For h1, the cumulative congestion window size remains larger than the window size of single h2 connection. The three pairs of box-plots under each network condition show the PLT distributions for webpages of size 2 MB, 8 MB, and 12 MB, respectively. …

[1]  Hari Balakrishnan,et al.  Stochastic Forecasts Achieve High Throughput and Low Delay over Cellular Networks , 2013, NSDI.

[2]  David Wetherall,et al.  How Speedy is SPDY? , 2014, NSDI.

[3]  Yiping Chen,et al.  Is HTTP/2 really faster than HTTP/1.1? , 2015, 2015 IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS).

[4]  K. Ramakrishnan,et al.  Towards a SPDY’ier Mobile Web? , 2015, IEEE/ACM Transactions on Networking.

[5]  Martin Thomson,et al.  Hypertext Transfer Protocol Version 2 (HTTP/2) , 2015, RFC.

[6]  Konstantina Papagiannaki,et al.  Is the Web HTTP/2 Yet? , 2016, PAM.

[7]  Moritz Steiner,et al.  Detecting Cellular Middleboxes Using Passive Measurement Techniques , 2016, PAM.