Research
From Jamjoom
Contents |
[edit] Network-Oriented Service Controls
My graduate research has primarily focused on developing network-oriented mechanisms to provide QoS support and overload protection of Internet services. It is a mixture of analytical modeling and empirical analysis of real systems and touches upon various aspects of network and software systems. This work demostrates how analytical modeling provides greater insight into the problem at hand, while empirical validation provides evidences for real-world deployments.
[edit] Persistent Dropping: A new Defense against Flash Crowds
Two working implementations of persistent dropping are designed and implemented that can be deployed in routers or end-servers. They are based on an analytical model that relates the drop probability and the connection-establishment delay, which can easily dominate the total client-perceived latency when accessing an Internet service. This model is constructed such that it can be applied to a wide range of drop policies. With this capability, the effects of different drop policies are explored and persistent dropping , a simple and novel mechanism that minimizes the client-perceived delay and the number of retransmitted requests while achieving the same control targets as traditional control mechanisms, is proposed. In particular, persistent dropping achieves three important goals: (1) it allows routers and end-servers to quickly converge to their control targets without sacrificing fairness, (2) it minimizes the portion of client delay that is attributed to the applied controls, and (3) it is both easily implementable and computationally tractable.
Because incoming requests to an Internet service can be inherently bursty, a new multi-stage filter is also introduced to perform accurate traffic prediction of future arrivals based on the applied control policy. Traffic prediction was based on the persistent client model and was augmented with a rejection mechanism to create the Abacus Filter (AF), a new traffic policing mechanism. AF complements the persistent dropping technique by better adapting to changing arrival patterns. The deriving principle is to maximize the number of admitted requests subject to the maximum acceptable network delay constraint. By modeling the time of the next TCP retransmissions, AFs estimate the amount of available future capacity and only admits portion of a burst that, with high probability, will eventually be admitted. This way, bursty traffic or a sudden jump in network load does not affect the overall delay of successful connections.
[edit] Adaptive Packet Filters
APF is a general architecture for integrating signaling information with dynamic control. This mechanism requires minimal OS modifications to facilitate its deployment and maximize its portability. It uses a simple and robust technique that can provide immediate reaction to varying load conditions. An APF implements its components as plug-in modules. This allows it to easily integrate new control filters (i.e., persistent dropping and abacus filters). The current implementation targets Layer-4 network devices and shows full overload protection as well as service differentiation.
[edit] Emulating Flash Crowds
Rigorous analysis of the effects of controls on incoming traffic to an Internet service is performed. This is a crucial step for designing and evaluating any network control mechanism. Based on this empirical study, the persistent client model is introduced to capture how client requests do not simply go away when dropped. Instead, there is a long-range dependency between the applied controls and future request arrivals. The proposed model is a departure from traditional approaches where client models are based on observed traffic aggregates (mostly through traffic traces). In contrast, this model is created after careful analysis of a wide range of individual clients accessing different web services. Combining the observations of different clients, a coherent picture is built detailing the expected behavior of the aggregate behavior of arriving requests.
With this new model, several unexpected aggregate behaviors in routers and end-servers are unfolded during different load scenarios. In particular, there was a surprising discovery of re -synchronization of dropped requests. This has a strong implication on the control of bursty request arrivals. Specifically, it shows that traditional traffic policing mechanisms (e.g., Random Early Drop and Token Bucket Filters) often underestimate the effects of the enforcement policy on the underlying client-perceived latency.
Because this work relies on actual implementation and real measurements, Eve---a scalable, extensible, and programmer-friendly tool that allows for fast generation of clients to test high -capacity Internet servers---is built. Eve is used to simulate thousands, or tens of thousands, of persistent clients with very high accuracy. It is a crucial component in the evaluation process as it integrates the proposed user-model with on-line measurements.
