Conventional NAS Cluster Technology Is Inadequate.

Conventional NAS clusters are unable to realize the full potential of a cohesive cluster due to several architectural problems such as:

NAS heads that are the sole repository of meta data:

Meta data represents the storage hierarchy of directories and the files they contain (the name-space). The problem here is that without these servers there is no way to access the files even though the file contents are still available on the cluster. It is like losing a map: the place you are looking for is still there but you can no longer find it.

The NAS heads are so critical that they are usually in HA configurations (i.e. they come in pairs). Besides the potential catastrophic failure of both HA servers there is the more common problem of synchronizing the HA servers' meta data. The failed HA server's meta data repository quickly becomes stale as it was updated with all the transactions that occurred during its failure. When the failed server comes back on-line, it must be reconciled with the active server before it can be in a position to take-over. Of course this is only an issue when the HA servers each have their own copy of the meta data. Things are even more risky if they share access to the same meta data repository as that then becomes a single point of failure.

NAS heads that broker between clients and the cluster:

In some NAS clusters files must be contained entirely within a single host as they cannot span multiple cluster members. There are several problems with this architecture. The NAS head must make a host selection based upon transient information at the time the file is created. Selecting a host based upon available capacity is difficult as there is no way to predict how large the file will grow and so whether the host selected will continue to be able to accommodate the file. It is also difficult to base the host selection upon performance because the least busy host at the moment may be quite busy soon.

These issues can lead to situations where an application fails to write data due to a lack of space on its host whereas the rest of the cluster is under-utilized. Similarly all the I/O for a file are directed to a single host while the rest of the cluster is not utilized. Performance can be further exacerbated when replication of the file across pairs of hosts is enabled as the write operations are often sequential (i.e. the primary host forwards the write request to the replication host).

Cluster members that do not cooperate:

NAS clusters that are not designed for cooperation between the members are only providing namespace virtualization over a group of independent storage islands. Realizing the full potential of a cluster is not possible without the members cooperating and helping each other. Such cooperation is needed to increase performance by sharing the workload across the cluster. A high level of cooperation is also needed to perform error correction both on-the-fly as well as part of healing of the cluster.

A Peer Fusion Cluster is a Tightly Integrated Group of Storage Peers.

A Peer Fusion NAS cluster is more than the sum of its parts. It is architected for strong performance and resiliency through cooperation between its peers:

The cluster owns the meta data:

The meta data in a Peer Fusion cluster is stored on every peer. This eliminates single points of failure and also assists in the healing process by ensuring that a quorum of peers agree on all data restored. A Peer Fusion NAS head is a gateway to the cluster to facilitate communication and has no persistent cluster meta data or file content. It has no data to reconcile upon failure which means maximum cluster availability as it can never be stale.

NAS heads that broker between clients and the cluster:

In a Peer Fusion cluster file data is striped across all the peers in the cluster which has many advantages: the aggregate available storage of the entire cluster is available to each file, so files can grow as large as the cluster and no file depends on access to a single peer. Another advantage is that all peers participate equally in I/O requests so the cluster load is inherently balanced, so performance is always optimal. Reading or writing a gigabyte file in a ten-peer cluster requires each peer to read or write one hundred megabytes only and this is done in parallel. The more peers are available in a Peer Fusion cluster the less load per peer so the faster the cluster.

Cluster members that do not cooperate:

The peers in a Peer Fusion NAS cluster are continuously cooperating. They cooperate in parallel to satisfy I/O requests, perform on-the-fly repairs, heal as well as restripe the cluster. This high level of cooperation ensures that the cluster is resilient yet it requires no additional communication thereby achieving a high performance.