QUIC performance in LSWS

[Boaz Reicher]

Hi everyone,
I am conducting an evaluation of QUIC, and I'm using LiteSpeed as the QUIC server.
I seem to get much lower throughput with LS, as opposed to the Chromium test server.
The following shows the throughput when downloading a 6 MB file multiple times in succession, using LSWS:



The following shows the throughput when downloading a 6 MB file multiple times in succession, using the Chromium test server:

The maximum available capacity (for both tests) is 100 Mbps.
For both tests I used the Chronet Android client. I also tries a couple of other QUIC clients, and the results are the same.

Both servers run on the same hardware and have the same network config.

I'm using the 5.2.3 release (2-CPU (LiteMage Unlimited)) with the trial license.

I'm trying to figure out why the Chromium test server (which is just a simple test implementation) seems to outperform LSWS' QUIC implementation.
Are there any settings I should look at?
Any help would be appreciated.

Thanks,

 

[Jon K]

Hi Boaz, is it possible to get a bit more information from you?

Were these tests run over a local environment or through the internet?
How did you monitor the download speed and generate the graphs?
What kind of setup did you run for client and server?
Is it possible to see your configurations?
What is the RTT?

 

[Boaz Reicher]

Hi Jon,
1. The tests were conducted over the internet.
2. The graphs where made using Wireshark captures, taken on the client. The download speed in the graph is the average per 50ms bin (Mbps = number of bytes captured in 50ms X (1000 / 50) X 8 / (1000 * 1000) )
3. The client is an Android application using the Chronet QUIC implementation (used other clients as well, such as Chrome from a desktop, got the same results).
As for the server setup, both the Chromium and Litespeed servers run on identical OpenStack VMs with 2 VCPUs and 4GB RAM and running Red Hat 7.1
4. Sure, what configurations do you want?

 

[Boaz Reicher]

Hi Jon,
Any ideas about this? I've also ran tests with simulated network impairments (bandwidth limit, packet loss, etc...) and found some interesting behaviors for the different QUIC implementations.
Are there any more configurable QUIC parameters except those exposed in the web UI?

Thanks,