Summer Camp 2011 Measurement Brainstorm
Clone this wiki locally
BISmark Summer Camp 2011/Measurement Brainstorm
- Wireless performance in the home
- Streaming performance
- Low-overhead bandwidth measurements
- Passive measurements (what kinds, and how?)
- Network neutrality / application performance monitoring
- Wireless mesh networking (e.g., getting bismark running on mesh potatoes, area deployments)
Inside the Home
- Wireless performance
- Throughput to wireless devices in the home
- Other wireless networks that exist in the "neighborhood" and how much traffic is being sent on these networks
- User behavior
- What does "normal" user behavior look like?
- Application behavior
- Automated vs. human-generated traffic (What does "ambient" traffic look like?)
- How much traffic is inside the home (internal: mDNS, bonjour, ARP, internal application traffic) vs. external?
- How much traffic is cacheable? (as a follow-up: how would caching any or all of this content improve performance?)
- How much traffic belongs to different applications? (e.g., is it true that 20% of traffic at peak time is NetFlix?)
- Device inventory and characterization/fingerprinting, including home automation devices
- Do people modify the router itself? (configuration changes?)
- What other information do applications leak? (monitoring set-top boxes for
data leaks, etc.)
- How much sensitive information is leaked in plaintext?
- Does the use of a particular application affect the performance of others (e.g., on the access link)?
- "HomeIDS" - perform active and passive measurements to detect anomalies (e.g., compromised hosts) inside the home
- Splitting snort/bro: generate signatures, but perform signature matching "in the cloud". Note: these signatures might have to be based on flow statistics if they are exported from the box.
- Factoring/locating bottlenecks: host, home network, access link, peering, server
On the Access Link
- Application performance (e.g., Skype, Netflix)
- Passive monitoring
- Active measurements (of statistics that are relevant to the application)
- Effects of NAT on performance
- Effects of split TCP on performance (Jacopo is working on this)
- Are the ISP policies static or dynamic? Within the access network, where is
each policy applied?
- Based on user behavior? Do the ISPs build profiles of users and treat each user's traffic differently?
- Based on destination? (e.g., do users behind the same head end experience different performance/policies than those in different locations?)
Topology and Applications
- What does the topology of the access network look like (e.g., by tracerouting
from Bismark node to Bismark node)?
- Mapping out sup-IP topologies in a region
- Which access ISPs peer with one another? Does peering vary by region?
- CDN Performance (Srikanth and Walter are working on this)
- What types of tricks do CDNs play in mapping clients to servers? What does DNS resolution look like from different geographies and providers?
- How does a CDN's peering affect the performance that users in different access ISPs see?
- Does fetching content from a peer yield better or worse performance than fetching from an origin?
- Quantifying, measuring, and monitoring "information bubbles" (see below)
(region-specific content, filtering, etc.).
- action item: add some wgets to bismark-active for different Web servers, somehow do "diffs", upload the results to a portal
- Combination of heavyweight and lightweight bandwidth tests to achieve more continuous measurements?
- Could the same set of measurements be applied on a mobile device?
Next Steps and Other Ideas
- Get TIE integrated with Bismark
- Identifying human-generated vs. automated traffic
- Identifying internal vs. external
- Kandula: Expose (learning communication patterns by grouping related flows)
This is a project for understanding and quantifying information filter bubbles (http://www.thefilterbubble.com), using a distributed measurement infrastructure like Bismark.
Filter bubbles focus on the impact of web personalization based on regional and personal characteristics. A user is exposed to things that he will probably like and be interested to. This creates a virtual environment which looks very familiar and friendly, removing "irrelevant" and/or "unpopular" content, where relevance and popularity is based on a rich vector of personal, regional, technical characteristics, and limits the breadth of user's experience.
The argument so far is on a relatively high-level. A more systematic and detailed study could help us understand and quantify how these reflects to the experience users perceive.
Using a distributed measurement platform like Bismark we can generate identical search requests from different regions to study such behaviors.
- Home Gateways act as clients. They issue search requests in multiple search engines (Bing, Google, Yahoo, - anything we could do with Facebook?), and report the results back to a central server.
- Requests should not be hardcoded in the image. We don't expect a query to have time variance, even if it has we can't predict it. The clients can fetch a queriy from the server every day and then issue the requests to the search engines (aka the "tip of the day")
- What sort of requests would reveal differences? Should it be up-to-date information like news, or static content?
- Get personalized behavior seems more difficult. It requires fake profiles, but how do you build history on this?
- What other things can we detect? Censorship, ...?
- Another way to do this would be using Tor and selecting the exit node.