Emc IP Holding Company LLC

Emc Corporation

Daniel J Sledz

John Michael Hayden

John Gemignani, Jr

Brett Johnson

Implementations are provided herein for establishing a failover cluster in a distributed file system that upon the occurrence of failover event, allows clients to actively migrate during ongoing file system activity from a source cluster to a target cluster without having to unmount an NFS export from the source cluster and remount the NFS export on the target cluster. Upon the occurrence of a failover event, clients can be rerouted to a target cluster that contains mirrored copies of the data the client was expecting to be stored on the original source cluster. However, in attempting to access the data, without unmounting and remounting an export, the client will continue to reference source cluster FSID and LIN identifiers when making NFS calls to the target cluster. Thus, implementations are provided herein for translating the requests of an NFS client on a target cluster after a failover event has occurred. A client can make NFS calls using FSID and LIN information of the original source cluster, and that information can be translated, e.g., through a reverse LIN map, to the unique LIN's of the target cluster, such that any failover migration of the client is transparent to the client.

NFS cluster failover

Implementations are provided herein for dynamically tuning resource allocation within a node among a cluster of nodes of a distributed file system. Each node can have a plurality of protocol heads that enable the node to communicate with clients using varying protocols. As tasks are received by the node from the client through the protocol head, the amount of work required to perform the task can be estimated, the amount of resources consumed by the task can be estimated, the amount of kernel memory required to perform the task can be estimated, and the amount of general memory required to perform the task can be estimated. Using these estimations, tasks can be scheduled by a scheduler based on the most efficient means to concurrently process as many tasks as possible using available resources of the node. In addition to efficiently scheduling tasks by the scheduler, a dynamic tuner can also reside on the node that can monitor all work executing on the node for each protocol head. The tuner can then distribute and, over time, redistribute resources to each protocol head using real-time data to make adjustments. In addition to tuning resources allocated to the protocol heads, the tuner can also account for pending and scheduled system task activity, as well as account for misbehaving clients.

Growing community of inventors

Seattle, WA, United States of America

Daniel J Sledz