Citation Context

...es (i.e. the IP addresses and port numbers) of responsible peers for id. Each peer maintains a routing table to forward requests it is not responsible for. Most DHTs, such as Chord, Pastry, or P-Grid =-=[11, 9, 1]-=- create routing tables of size O(log(n)), where n is the number of peers in the system. The routing entries are chosen in such a way that the resulting graph has small-world properties [5, 4]. Resolvi...

Citation Context

...es (i.e. the IP addresses and port numbers) of responsible peers for id. Each peer maintains a routing table to forward requests it is not responsible for. Most DHTs, such as Chord, Pastry, or P-Grid =-=[11, 9, 1]-=- create routing tables of size O(log(n)), where n is the number of peers in the system. The routing entries are chosen in such a way that the resulting graph has small-world properties [5, 4]. Resolvi...

Citation Context

...Grid [11, 9, 1] create routing tables of size O(log(n)), where n is the number of peers in the system. The routing entries are chosen in such a way that the resulting graph has small-world properties =-=[5, 4]-=-. Resolving any lookup request started at any peer is then guaranteed to take O(log(n)) overlay hops on average. There are two ways of resolving lookup requests in a DHT: when recursive routing is app...

Citation Context

...w papers have discussed CC in DHTs so far. For recursive routing many DHT proposals use either TCP for neighbor-to-neighbor routing or a simplified reimplementation of TCP using UDP, e.g. in Tapestry =-=[14]-=-. However, it is not further specified whether these systems contain DHT-wide CC mechanisms and how they behave under heavy loads. Most of the current DHT implementations are designed for low to moder...

Citation Context

...es (i.e. the IP addresses and port numbers) of responsible peers for id. Each peer maintains a routing table to forward requests it is not responsible for. Most DHTs, such as Chord, Pastry, or P-Grid =-=[11, 9, 1]-=- create routing tables of size O(log(n)), where n is the number of peers in the system. The routing entries are chosen in such a way that the resulting graph has small-world properties [5, 4]. Resolvi...

Citation Context

...CC mechanisms and how they behave under heavy loads. Most of the current DHT implementations are designed for low to moderate loads and CC has not been an issue yet. For iterative routing, in DHash++ =-=[3]-=- presents the Striped Transport Protocol (STP), which has similarities with CSCC as it maintains DHT-wide destinationindependent congestion state. The difficulty in iterative routing, however, is that...

Citation Context

...eptably most of the time. In the past years, however, the usage of DHTs has passed simple file sharing. A large research community is now working on providing full-text retrieval on top of DHTs, e.g. =-=[6, 12]-=-. Performing information retrieval on top of a DHT (P2P-IR) drastically increases the load of the system by at least two orders of magnitude, compared to file sharing. In P2P-IR each peer extracts sev...

Citation Context

...tion-specific optimizations regarding (among other things) congestion control in peer-to-peer networks. CC in DHTs is orthogonal to most of the load balancing work that has been done for P2P systems. =-=[8]-=-, for example, proposes mechanisms to reduce the imbalance of data items stored at peers. Though important for DHTs, these mechanisms cannot avoid congestion caused by routing traffic in a DHT. Furthe...

Citation Context

...eptably most of the time. In the past years, however, the usage of DHTs has passed simple file sharing. A large research community is now working on providing full-text retrieval on top of DHTs, e.g. =-=[6, 12]-=-. Performing information retrieval on top of a DHT (P2P-IR) drastically increases the load of the system by at least two orders of magnitude, compared to file sharing. In P2P-IR each peer extracts sev...

Citation Context

...mission timeouts. However, when the network becomes congested, latency predictions might not be accurate anymore. Furthermore, it becomes difficult to distinguish between failed and overloaded peers. =-=[2]-=- proposes component-based transport protocols for highly distributed applications. They provide higher flexibility to perform application-specific optimizations regarding (among other things) congesti...

Citation Context

...t and resolve. 2.2.2 Deadlock-Free Back-Pressure A known principle in packet-switched networks is that a routing scheme is deadlock free if and only if there are no cycles in the buffer waiting graph =-=[10, 7]-=-. There have been many proposals for routing in DHTs, e.g. [11, 9, 1], many of which build a ring or a tree topology. We will show in the following that for ring and tree topologies it is possible to ...

Citation Context

...Grid [11, 9, 1] create routing tables of size O(log(n)), where n is the number of peers in the system. The routing entries are chosen in such a way that the resulting graph has small-world properties =-=[5, 4]-=-. Resolving any lookup request started at any peer is then guaranteed to take O(log(n)) overlay hops on average. There are two ways of resolving lookup requests in a DHT: when recursive routing is app...

Citation Context

...ta items stored at peers. Though important for DHTs, these mechanisms cannot avoid congestion caused by routing traffic in a DHT. Further work on P2P CC is in the area of multi-cast overlay networks. =-=[13]-=-, for example, studies the performance of a back-pressure mechanism. 5 Conclusions In this paper, we have introduced congestion control for DHTs. To the best of our knowledge, this is the first paper ...

Citation Context

...t and resolve. 2.2.2 Deadlock-Free Back-Pressure A known principle in packet-switched networks is that a routing scheme is deadlock free if and only if there are no cycles in the buffer waiting graph =-=[10, 7]-=-. There have been many proposals for routing in DHTs, e.g. [11, 9, 1], many of which build a ring or a tree topology. We will show in the following that for ring and tree topologies it is possible to ...