Proceedings Template word
Download 493.43 Kb. View original pdf
|
| Map to server. Once assigned to a specific cluster, a low- level map within the cluster directs the user to a specific machine, based on factors including the likelihood of that machine to have the requested piece of content in cache. It is desirable to maintain locality within clusters (mapping requests for the same piece of content to the same machine, as this improves performance and makes efficient use of cache space. The challenge is to do so in a dynamic environment that also factors in load changes and machine failures. Akamai‘s pioneering research in this area began with consistent hashing in [21] and [22] over a decade ago, and has evolved significantly from that point. When hardware and network faults are identified (such as a failing hard drive on a server, the failed edge servers are suspended and will finish up processing in-progress requests but will not be sent any additional end users until the fault has been resolved. A more detailed example is provided in Section 8. The mapping system itself is a fault-tolerant distributed platform that is able to survive failures of multiple data centers without interruption. The scoring and map-to-cluster portions of the system run in multiple independent sites and leader-elect based on the current health status of each site. The map-to-server portions of the system run on multiple machines within each target cluster. All portions of the system, including monitoring agents and DNS servers, communicate via a multi-path overlay transport described in more detail in Section 7.3) that is able to tolerate network faults. Because the efficacy of the mapping system is important to the overall performance of Akamai‘s system, there is continuous research and development being performed to advance and refine it. This includes ongoing work to improve the quality of scoring data inputs, improve locality for large-footprint content, address shifts in Internet architecture, develop new methods for enhanced fault-isolation, and optimize the performance of system components to enable even faster response times. One example of such a refinement is work that was done to enhance the servers abilities to adjust their own capacity inputs to the mapping system based on self-monitoring of resource utilization. This enables heterogeneous hardware, running different types of applications, to be utilized more efficiently throughout the network. The sheer volume (and frequency) of data being processed in the mapping system also presents a challenge. During an early system redesign, analysis of the initially proposed design indicated that the amount of data that would need to be collected and communicated would have exceeded the amount of end user traffic being served. Much work has been done since, and continues to be done, to reduce this data communication burden while retaining all information essential to high-quality mapping. Download 493.43 Kb. Share with your friends:
|