The freshman handbook: a hint for the server placement of social networks

Ying Zhang, Ying Zhang 0022, Mallik TatipamulaWWW 2011 · DOI

TL;DR

New social networks face a critical challenge: where to place their servers to minimize latency and data sharing costs. This paper proposes a novel approach for these "freshman" OSNs, leveraging public data from established social networks to intelligently predict optimal server locations, even without historical user data.

What Problem Does This Paper Address?

Launching a new Online Social Network (OSN) comes with a crucial, early decision: where to deploy servers. This isn't just about having enough capacity, but about strategic placement to ensure fast response times for users and efficient data exchange between servers. Traditional methods for server placement often rely on analyzing historical user request data from existing servers. However, for a new OSN, this historical data simply doesn't exist. This paper tackles this "cold start" problem, aiming to provide intelligent server placement suggestions for nascent social networks by analyzing publicly available information from established OSNs.

Methods

The authors propose a two-pronged approach:

  1. Scalable Server Placement Algorithms: They introduce algorithms designed to optimize server placement. One method uses graph partitioning to cluster users based on their geographic locations, aiming to minimize the distance between users and their assigned servers. Another algorithm specifically considers inter-user data sharing, which is a significant factor in social networks due to strong community structures.
  2. Leveraging Public OSN Data: The core innovation lies in using data from existing, popular OSNs (like Amazon reviews, Buzz, Twitter, and Flickr) to inform placement decisions for new OSNs. They analyze user distribution and friendship patterns across these established platforms to identify commonalities and derive generalizable insights. This "freshman handbook" approach allows new OSNs to benefit from the collective experience of older ones.

Datasets

The paper utilizes data from four popular online social networks:

  • Amazon review system: Likely used to infer user distribution and potentially social connections through review patterns.
  • Buzz: A social networking service from Google.
  • Twitter: A microblogging and social networking service.
  • Flickr: A photo-sharing platform.

The data consists of user profiles and their associated friend lists. The authors mention that while ideal data would be direct latency measurements, they approximate this using hypothetical direct link latency based on geographic locations extracted from user profiles.

Key Insights and Ideas

  • Locality of User Distribution: Despite the global nature of the internet, users of OSNs tend to concentrate in specific geographic locations (e.g., metropolitan areas). This is a fundamental principle that can guide server placement.
  • Commonality in Server Placement: Surprisingly, even with different application types (e.g., microblogging vs. photo sharing), popular OSNs exhibit significant commonalities in their chosen server deployment locations. This suggests that generalizable patterns exist.
  • Friendship Distribution Matters: The geographic spread of a user's friends can significantly influence server placement decisions, especially when considering inter-server communication costs. Platforms with more geographically dispersed friend networks (like Flickr) might require different placement strategies than those with more localized connections.
  • Public Data is a Goldmine: The paper demonstrates that publicly accessible data from established OSNs can be a valuable resource for new entrants, providing a shortcut to intelligent infrastructure planning.

Key Results

The study found that:

  • A small fraction of geographic locations host the majority of users across the analyzed OSNs. For instance, 70% of Flickr users were found in just 5% of locations.
  • When deploying 10 servers, the locations chosen by the algorithms for Buzz and Twitter were remarkably similar, with 8 out of 10 locations overlapping.
  • Amazon's distribution showed slight differences, attributed to its distinct application characteristics.
  • Flickr exhibited a more spread-out friendship distribution compared to Buzz and Twitter, indicating that platforms catering to niche interests might have different connectivity patterns.

Limitations and Future Work

  • Approximation of Latency: The paper acknowledges that using hypothetical direct link latency is an approximation and not a direct measurement of real-world network latency. Actual network conditions can be more complex.
  • Data Availability and Privacy: While the paper leverages public data, the extent and granularity of this data can vary, and privacy concerns might limit access to the most detailed information.
  • Dynamic Nature of OSNs: Social networks are constantly evolving. The proposed methods are based on snapshots of existing OSNs, and future work could explore how to adapt placement strategies to dynamic user growth and changing network topologies.
  • Joint Analysis Approaches: The paper mentions exploring two approaches for joint analysis (after individual OSN selection vs. incorporating multiple OSNs upfront) but doesn't definitively state which is superior, leaving room for further investigation.

Summary updated

BibTeX
@inproceedings{zhang2011freshman,
  author = {Ying Zhang and Ying Zhang and Mallik Tatipamula},
  title = {{The freshman handbook: a hint for the server placement of social networks}},
  booktitle = {The Web Conference},
  year = {2011},
  doi = {10.1145/1963192.1963280},
}
Figure