Graceful Degradation – Georgia Tech – Advanced Operating Systems

Graceful Degradation – Georgia Tech – Advanced Operating Systems


And the DQ principle is also very useful for managing graceful degradation of service. So DQ defines to total system capacity. So if a server is saturated, meaning that we have reached the limit of the server in terms of DQ. That’s a constant. DQ’s a constant. And so if you reach that limit, then we have a choice of graceful degrading the service from the point of view of the client. One possibility is we keep the harvest the same meaning that every client request that comes in has complete fidelity in terms of the answers returned by the server. So D is fixed. Q comes down because DQ is a constant. That’s one option. The other option, of course, is to keep the volume of clients that are service, that is the yield Q to be a constant, but decrease the harvest. So the fidelity of the results returned to the users is less than. 100%, but we keeping more of the user community happy by serving more of them. Because DQ is a constant, it allows us to gracefully degrade the service being provided by the server depending on the choice you want to make in terms of fidelity of the result or the yield that you want to provide the user community. In other words. The DQ principle gives us an explicit strategy for managing saturation. So as a system administrator, when we make these decisions on which to sacrifice and which to keep constant, we know how our decisions are going to affect the harvest, how it’s going to affect the yield. How it is going to affect the up time and so on. So the choices that a system provider has or strategies that a system provider can use in structuring the servers, knowing that DQ is a constant, is when the server is saturated. They can do cost based admission control. You pay more, you get more. That may be one way to do it. Or priority or value based admission control. That may be another way to deal with service saturation or reduce data freshness. That is the harvest may be the same, but you reduce the fidelity of the data. For instance, if I’m serving videos, and I can serve the videos at different bit rates, if the server is saturated, I might decide to serve the video to all the users, all the videos that they want, but at a lower bit rate. So that is reducing the fidelity. Of the data that is being harvested. So that’s the idea behind the DQ principle, how it might be used for graceful degradation of service when the server is saturated.