In-Depth Investigation of CDN: Part 2

In this lesson, we learn how content consistency can be achieved using different consistency mechanisms. We also learn about where we should deploy the proxy servers and the difference between CDN as a service and specialized CDN.

Content consistency in CDN

Data in the proxy servers should be consistent with data in the origin servers. There’s always a risk of users accessing stale data if the proxy servers don’t remain consistent with the origin servers. Different consistency mechanisms can be used to ensure consistency of data, depending on the push or pull model.

Periodic polling

Using the pull model, proxy servers request the origin server periodically for updated data and change the content in the cache accordingly. When content changes infrequently, the polling approach consumes unnecessary bandwidth. Periodic polling uses time-to-refresh (TTR) to adjust the time period for requesting updated data from the origin servers.

Time-to-live (TTL)

Because of the TTR, the proxy servers may uselessly request the origin servers for updated data. A better approach that could be employed to reduce the frequency of refresh messages is the time-to-live (TTL) approach. In this approach, each object has a TTL attribute assigned to it by the origin server. The TTL defines the expiration time of the content. The proxy servers serve the same data version to the users until that content expires. Upon expiration, the proxy server checks for an update with the origin server. If the data is changed, it gets the updated data from the origin server and then responds to the user’s requests with the updated data. Otherwise, it keeps the same data with an updated expiration time from the origin servers.

Leases

The origin server grants a lease to the data sent to a proxy server using this technique. The lease denotes the time interval for which the origin server agrees to notify the proxy server if there’s any change in the data. The proxy server must send a message requesting a lease renewal after the expiration of the lease. The lease method helps to reduce the number of messages exchanged between the proxy and origin server. Additionally, the lease duration can be optimized dynamically according to the observed load on the proxy servers. This technique is referred to as an adaptive lease.

In the following section, we discuss where to place the CDN proxy server to transmit data effectively.

Deployment

We have to be clear with the answers to the following questions before we install the CDN facility:

• What are the best locations to install proxy servers to maximally utilize CDN technology?

• How many CDN proxy servers should we install?

Placement of CDN proxy servers

The CDN proxy servers must be placed at network locations with good connectivity. See the options below:

• On-premises represents a smaller data center that could be placed near major (IXPs)An internet exchange point (IXP) is a physical infrastructure where multiple networks connect and exchange internet traffic directly to improve speed and reduce costs..

• Off-premises represents placing CDN proxy servers in ISPsAn internet service provider (ISP) is a company that provides individuals and organizations access to the internet and related services.’ networks.

Today, it might be feasible to keep a large portion of a movie’s data in a CDN infrastructure that’s housed inside an ISP. Still, for services like YouTube, data is so large and ever-expanding that it’s challenging to decide what we should put near a user. Google uses split TCP to reduce user-perceived delays by keeping persistent connections with huge TCP windows from the IXP-level infrastructure to their primary data centers. The client’s TCP requests terminate at the IXP-level infrastructure and are then forwarded on already established, low latency TCP connections.

Doing this substantially reduces client-perceived latency, which is due to the avoidance of the initial three-way handshake of TCP connection and slow-start stages to a host far away (had the client wanted to go to the primary data centers of Google). A round-trip delay to IXP is often very low. Therefore, three-way handshakes and slow starts at that level are negligible. Predictive push is a significant research field to decide what to push near the customers.

We can use measurements to facilitate the decision of proxy server placement. One such tool is ProxyTellerA proxy placement tool for content delivery under performance constraints. Source: Trianfillou, P. and Aekaterinidis, I., 2003, December. Proceedings of the Fourth International Conference on Web Information Systems Engineering, 2003. WISE 2003. (pp. 62-71). IEEE. to decide where to place the proxy server and how many proxy servers are required to achieve high performance. ProxyTeller uses hit ratio, network bandwidth, and client-response time (latency) as performance parameters to decide the placement of proxy servers. Other greedy, random, and hotspot algorithms are also used for proxy server placements.

Note: Akamai and Netflix popularized the idea of keeping their CDN proxy servers inside the client’s ISPs. For many clients of Akamai, content is just one network hop away. On the other hand, Google also has its private CDN infrastructure but relies more on its servers near IXPs. One reason for this could be the sheer amount of data and the change patterns.