Re: H2 Server Push performance data

Hi Tom,

Thanks for your feedback!

The first couple of charts (based on the multiple customer sites) use the mean as a way to provide a summary view of performance for each individual site.  The last chart in the article is a percentile plot for one customer site which demonstrates a more granular view, but we do not have percentile plots put together for each customer site. It should be possible to do that in the event we were to perform deeper analysis on more sites.

The A/B data was not collected in a way that we would consider to be pure enough to be considered i.i.d coin flips. There exist dimensions where the values can make it more or less likely that Server Push may be applied on specific responses within the production network (e.g. geo, time of day, user-agent), and so the approach was chosen to use linear regression to attempt to control for these variables.

I agree that the FirstContentfulPaint metric would be useful to analyze; we had some limitations in our dataset regarding collection of that metric that resulted in it not being reliable enough to use (it wasn't as widely collected as DOMComplete). There have been some encouraging results seen for this metric based on some preliminary analysis, but we wouldn't consider the data and analysis reliable enough yet for sharing.

Thanks,
Aman


________________________________
From: Tom Bergan <tombergan@chromium.org>
Sent: Monday, July 22, 2019 2:02 PM
To: Nanner, Aman
Cc: ietf-http-wg@w3.org
Subject: Re: H2 Server Push performance data

Hi Aman,

Interesting results, thanks for sharing! I do wonder why you decided to compare means instead of percentiles? It looks like there are some crazy bad outliers. Assuming your "April 2019 Results with Old Methodology" and "April 2019 Results (1% quantile excluded)" graphs use the same set of sites in the same order, the mean goes from about +500ms to -200ms. If 99% of requests on that site take -200ms, the last 1% must take 70s! Instead of means, you could have compared medians, or 25th/75th percentiles, or other percentiles in the distribution. You can compute confidence intervals on percentiles.. You can also do something like Mann-Whitney to check if the distributions differ significantly.

I'd also like to hear more about this linear regression that you run on the A/B results. I couldn't follow why this regression is necessary. If your A/B test uses i.i.d. coin flips, you can compute confidence intervals directly on the A/B results. Is your A/B test not i.i.d.? How do you know the linear regression doesn't introduce bias?

Also, have you looked into measuring other metrics, such as FirstContentfulPaint?

-Tom

On Mon, Jul 22, 2019 at 4:41 AM Nanner, Aman <ananner@akamai.com<mailto:ananner@akamai.com>> wrote:
At the previous IETF 102 HTTPWG session in Montreal, I had presented some Akamai data on H2 Server Push which can be found here:

https://github.com/httpwg/wg-materials/blob/gh-pages/ietf102/akamai-server-push.pdf<https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_httpwg_wg-2Dmaterials_blob_gh-2Dpages_ietf102_akamai-2Dserver-2Dpush.pdf&d=DwMFaQ&c=96ZbZZcaMF4w0F4jpN6LZg&r=HVAbQ4bnkmcYKMR6_H0nEMfpVIpY1jgJWVBhC_es5Nw&m=JhnrDdJ2FDK0-S_OawS0GoKyXIVn4C8H_c03AveZDRo&s=UzEaCDDavuWPtu3zi_hdJa51YwKMwdzescfn9rpYnlQ&e=>

Akamai has conducted some more recent tests with a tweaked methodology (exclusion of 1% highest-latency requests on the long-tail), and we have found some interesting results. I share some more details about the performance analysis here:

https://medium.com/@ananner/http-2-server-push-performance-a-further-akamai-case-study-7a17573a3317<https://urldefense.proofpoint.com/v2/url?u=https-3A__medium.com_-40ananner_http-2D2-2Dserver-2Dpush-2Dperformance-2Da-2Dfurther-2Dakamai-2Dcase-2Dstudy-2D7a17573a3317&d=DwMFaQ&c=96ZbZZcaMF4w0F4jpN6LZg&r=HVAbQ4bnkmcYKMR6_H0nEMfpVIpY1jgJWVBhC_es5Nw&m=JhnrDdJ2FDK0-S_OawS0GoKyXIVn4C8H_c03AveZDRo&s=5ScQ7t8fvPbCcHySFH1WDh9iHpnjchASXvAmdo8mvJM&e=>

Thanks,
Aman Nanner
Akamai Technologies Inc.