 How good or how bad could a content delivery network perform? well That could be determined at the very basic level using a mathematical model True. However, if you are interested in assessing the actual performance of large-scale content delivery networks, then we may have to carry out experimental Analysis for that in this module. We are going to look at a very well-known example from Raj Kumar Bhaiya in the content delivery network Book that I have already shared with you in this particular experiment Overlay networking that is overly routing which is the routing between Peering agents which does not involve the shortest paths is Is carried out and is compared to the direct or the shortest path Now this experimental setup would be then used for creating some data and that particular data is analyzed and some findings are made So what is the experimental setup? The experimental setup is to collect the data to optimize the shortest paths and the Indirect paths which are based on overlay Between edge networks. Now what are edge networks? Edge networks are tier 3 or tier 4 networks Which are residing either on the customer end or at the enterprise origin servers Now these particular edge servers are going to provision the objects which are requested by the client Now this could be one direct delivery mechanism like shortest path routing or BGP routing but sometimes because of the Assumption that everyone's going to take the shortest route the shortest path or the direct route turns out to be more Delay oriented for that routing overlays are basically specialized devices Which are routers or gateways Interior one networks which behave as an intermediate node for creating an indirect path between the server and the client The example platform that the authors have adopted is from Akamai Akamai is a well-known Content delivery network platform. So the measurements are done on CDN clusters the clusters are basically groups of servers Which are spread over the whole world now the Arrangement or the platform that the Authors have adopted is 1100 clusters from over six continents and 77 countries and you can see 630 cities. There are 1100 cut clusters Now these clusters are some of these are placed near the core servers around 15% of them and Around 85% of them are near the edge servers the intermediate nodes actually forward the traffic from the core servers because mostly the the Edge servers or the Origin servers are placed on the core ISP. So they use overlay routing to deliver the content to the clients Now Typically, there's an important point that typically these the servers in one cluster Are connected to a single service provider? It's something which is pretty obvious But it was declared as one of the design principles for this particular experimental study the data which was collected was collected using simple programming of Automation of pinging so it's an IPM ICM P echo request echo reply mechanism where 64 bytes of Requests were initiated from every Cluster to every other cluster so you can think about Mesh of 1100 clusters So it was done every two minutes so correspondingly since the number of nodes was very large So around the ten requests were being received every second at every node This activity was performed for four weeks and the duration was 1.5 hours Every day and the each path actually between every node to every every cluster to every other cluster was probed for 3,780 times it means that this was repeated so many times to make sure that the readings Are more average and there are hardly any anomalies. So the total probes turned out to be 652 million. It's a very large number. So this actually is a good Reliable data that can be banked upon so the overall findings were in the represented in the form of some tuples that included the time stamping identifiers of source and destination the direct round trip time the one hop shortest round trip time and Three hop shortest round trip time three hearts hop shortest round trip time actually means that the Indirect parts actually Instead of making direct trip go through up to three hops So this actually is the operation of how the overlay networking works This is more into the arrangement between different Edge and core serve core clusters. These clusters were placed across different continents seven continents and 15% of these were the core clusters and around 85% were the edge clusters So the mnemonics or the abbreviations are obvious for Asia. It is a s for South America. It is essay Now the findings actually start with the data collection So you you can see we have data which is categorized as Latency reduction percentage. What is that? It is actually the difference between the latency incurred in direct delivery of the object from the server to the client and The overlay path that is one two three pops Based round trip times now the difference between shortest path. That is the direct path and the overlay based path Actually was represented in the form of these five columns So the first category is actually the autonomous system to autonomous system That is you can say a cluster to cluster category. For example, Africa to Asia Now less than minus 10% actually means that the indirect path is in fact slower than the direct path That is against the intuitive argument that was being made Then we have 10% which is comparable almost the direct and indirect path would incur the same delay And then we have marginal improvement that is 30% improvement in terms of time using Indirect path or using overlay and then we have 30 to 50% significant improvement and then we have The more than 50% that is the indirect path takes Less than half of total time then it takes through the direct path Now the analysis between Of comparison between direct and indirect path is going to help us evaluate the performance And then we can think about making some recommendations But for analysis the data set was divided into buckets that I've already explained The findings actually are that the best indirect paths were 10% slower than the direct path now interestingly around 15% of the Paths between different clusters fell into this category Now was this argument Wrong that indirect paths result into better latency. Well, it is only 15% It means around 85% are the ones that have shown overall improvement See we have the 10% bucket where the direct and indirect are almost equal So 24.7% actually result into almost equal performance than direct path So it means that we can think about 40% overall scenarios or Paths where the autonomous system to autonomous system traffic was Being delivered almost equal to or less than equal to the direct path but overall the Results are quite encouraging. You can see for 23.4% the overall Improvement in indirect path is marginal. That is 30% up to 30% Then up to 50% is 13% paths then the Double reduction in latency is for 23% of the paths. This is a significant improvement So it means the overall argument that overlay networking results into better Utilization in terms of throughput and latency Is found to be correct now This is I just wanted to share with you that look at it again and you can make Even more interesting insights