Citation Context

...ification to hard state (such as changes to a file in the Live Mesh synchronization service) and user-visible soft state (such as device connectivity information stored on a pool of in-memory servers =-=[1]-=-). The Live Messenger trace contains all login and logoff events, all IM conversations and the participants in each conversation, and the total number of messages in each conversation. The Live Messen...

Citation Context

...date any extant sequencers before starting work, so that no remnant writes can succeed thereafter. The sequencer is functioning as a lease enforcement mechanism, in a similar manner to the Centrifuge =-=[3]-=- system. Thus, we can guarantee that, for a given key, only a single worker can write to that key at a particular point in time. This single-writer guarantee is also critical to the maintenance of sof...

Citation Context

...tion requires that all updates to a given partition key are being sent to the same server. This requirement could be met using a variety of techniques; the applications we evaluate rely on Centrifuge =-=[2]-=-. Centrifuge is a system that combines leasemanagement with partitioning. Centrifuge uses a logically centralized manager to divide up a flat namespace of keys across the middle-tier servers. Centrifu...

Citation Context

...model has led to its adoption in a variety of enterprise, datacenter and wide-area network environments [1,2,3,4]. Microsoft, Google and Yahoo, for instance, usepub/sub forend-usernotificationdelivery=-=[5]-=-, data disseminationin large-scale server farms [3], and in distributed data storage systems [2]. In enterprise settings, pub/sub has appeared in severalcontexts and standards including EnterpriseServ...

Citation Context

...= ri(t)/t0+ ui/maxa≤ j≤n{u j}. In order to build up a scheduling objective with the hit rate and the usage ratio, we define a composite cost from hit rate and usage rate as c=α ·h+(1−α) ·r, where α ∈ =-=[0,1]-=- is the impact factor to control which feature is more important and h = 1−h is a miss rate, and state the scheduling objective as follows: minimize ∑ni=1 (α ·hi+(1−α) · ri) subject to hi ∈ [0,1] and ...

Citation Context

...t preemption in Mapreduce [6], w.r.t. fairness [26], at intelligent killing of tasks [10] (including the Hadoop Fair Scheduler [23]), and SLOs (Service Level Objectives) in generic cluster management =-=[1, 30, 43]-=-. In comparison our work is the first to study the effect of eviction policies and deadline-based scheduling for resource-constrained Mapreduce clusters. Our strategies can be applied orthogonally in ...

Citation Context

... due to memory management. 5.2 Reconfiguration The reconfiguration protocol moves a FaRM instance from one configuration to the next. Using one-sided RDMA operations is important to achieve good performance but it imposes new requirements on the reconfiguration protocol. For example, a common technique to achieve consistency is to use leases [18]: servers check if they hold a lease for an object before replying to requests to access the object. If a server is evicted from the configuration, the system guarantees that the objects it stores cannot be mutated until after its lease expires (e.g., [7]). FaRM uses this technique when servicing requests from external clients that communicate with the system using messages. But since machines in the FaRM configuration read objects using RDMA reads without involving the remote CPU, the server’s CPU cannot check if it holds the lease. Current NIC hardware does not support leases and it is unclear if it will in the future. We solve this problem by implementing precise membership [10]. After a failure, all machines in a new configuration must agree on its membership before allowing object mutations. This allows FaRM to perform the check at the cl...

Citation Context

...s remain in synchrony by using the Paxos protocol to agree on the order of method invocations. There has been much work on employing the Paxos protocol to achieve fault-tolerance in specific settings =-=[34, 36, 4, 32, 10, 20, 1]-=-, in which Paxos was monolithically integrated into a specific, static API offered by the system. In contrast, OpenReplica is a general-purpose, open service that enables any object to be made fault-t...

Citation Context

...a cloud provider executes maintanance operations on the server cloud, e.g., rollout of security patch operations, that generates a continuous and unpredictable restarts of servers that can take hours =-=[1]-=-. Therefore, a rollout operation translates into servers that join and leave the storage service (i.e., server churn). As a consequence, a correct and highly available storage has to be ready to auton...

Citation Context

...ver is next in the hash space, as in distributed networks such as Chord [19]. Instead, each server is assigned a fixed section of the address space which is proportional to its capacity. Some systems =-=[6, 1]-=- which use consistent addressing mitigate the load balancing problems of Chord by mapping to many virtual servers using a hash function and then mapping sets of virtual servers to physical servers acc...