![[logo]](/tt.gif "[logo]")
TinyTube Blog TinyTube BlogPosted in: Uncategorized
After we removed a TinyTube bottleneck earlier this week by adding more servers, we noticed some of the pages were still loading slowly, particularly the search pages. After profiling the search appliance (more on that another day), it turned out it didn’t account for all of the page load latency. So we did more profiling.
As it turns out, ad pulls from various mobile ad networks account for a significant fraction of the page load time. If you’re not aware, the TinyTube ads are done through server-side include (SSI). It’s different from say, Google Adsense, for which the browser can load ads in parallel with the page. SSI requires all the ads to be in the page before it can leave the server. Search engine results pages (SERP) are the bulk of TinyTube page views, and thus contain the bulk of the ads.
I wanted to get some hard numbers on how much each ad network is slowing us down. So we collected data for each network, sample size is 100. Here’s a boxplot summarizing those measurements.
This is fractional seconds in log space (think earthquake/pH scale), so -1 compared to 0 means 1/10 the time.
For the search engines, the participants are, in alphabetical order:
There is a huge difference in the quality of service (QOS). Looks like Admob is a clear winner for speed. Decktrade and Millennial are clear winners for consistency, everyone else gives a > 1s response at some point.
However, there is a caveat to that interpretation, namely that these measurements were made from our California Datacenter, and are certainly affected by network proximity to the ad server. However, that’s not enough to explain the similar values for the other four given that Admoda is serving out of Europe. Some of the difference could also be accounted for by HTTP payload size. Millennial and DeckTrade are the same company, but Decktrade sends only text ads, whereas Millennial sends banner ads. The text ad payload is a lot smaller, so we would expect it to be marginally smaller and we do.
I’ll add some more data here when I get a chance. It will be interesting to see how adding the Quattro and MoPhap measurements might change interpretation, if at all.
The whole purpose of this exercise isn’t to point out which ad networks are better or worse, just that they are variable in terms of quality of service for TinyTube in particular. And that’s what it’s all about: getting fast page loads to raise the bar for our site.
Update: Raj Singh at Emerging Telephony has suggested polling the ad networks from several NOCs to see if the latency has a large geo-based component. Putting together a set of “test” publisher accounts to distribute to other mobile site operators around the globe is a great idea.
Interesting study. Maybe people looking to monetize their mobile websites should take this into consideration!
by Rico on 2007-08-09 at 9:16 am
I’ve read all the instructions and then what I want to know is that How can I get the full video clip from your web site and the way operation cost will be considered. Please help me.
by Lucas on 2007-08-09 at 1:09 pm