Discover Technology: EMC - iSilon

1. EMC Isilon Scale-out Storage Platform architecture

The EMC Isilon Storage Platform consists of a number of storage nodes that provide storage and network connectivity for the cluster, an internal Infiniband network and external networking connectivity that provides access to clients.

EMC Isilon operates as a scalable appliance behind a customer’s Ethernet infrastructure.

EMC Isilon uses standard 10 gigabit and one gigabit Ethernet and speaks industry standard protocols to provide a high level of compatibility with existing and new infrastructures.

1.1 Interconnect
The Cluster interconnect is InfiniBand, which was designed as a high-speed interconnect for high performance computing. The architecture can run on any suitably fast technology, InfiniBand just presents the best performance characteristics today.
The interconnect is very important to OneFS. It carries block I/O and internal message passing, so it needs to be both high throughput and very low latency. We are usually shipping the fastest standard that’s in mass production, and InfiniBand has never been a bottleneck.
A single front-end operation like a chmod can generate multiple messages on the back-end. There’s more to it than this, and frequently the lock will be cached. But the main point is that the reliability and performance of the interconnect is very important to creating a high-performance, true scale-out storage system.

1.2 External Network

The external networking components of an EMC Isilon cluster provide client access over a variety of protocols, including NFS, SMB, HDFS (Hadoop File System), FTP and HTTP. Each storage node connects to one or more external Ethernet networks using 1 or 10 Gigabit connections, or a combination of both. Each node supports either four (4) 1 Gigabit Ethernet connections, or two (2) 1 Gigabit and two (2) 10 Gigabit Ethernet connections. 1 Gigabit ports use standard RJ45 connectors and category 5e or 6 cables, while 10 Gigabit ports use either SFP+ optical or copper transceivers

1.3 Protection & Availability

OneFS supports very flexible data protection methods as granularly as an individual file or folder.

Data can be mirrored from two ways up to eight ways.

For data protection with lower overhead Forward Error Correction or FEC is used to distribute data in a way that is statistically likely to provide fault tolerance.

These data protection policies can be changed at any time and the data will be re-protected at its new level in the background.

FEC can support from one up to four simultaneous failures without losing access to data.

The nomenclature used here is N+M where N is the number of nodes containing data and M is the number of nodes or node components that can fail simultaneously.

Minimum cluster sizes are required to support higher p

rotection levels since quorum should be preserved in the event of tolerable simultaneous component failure.

In addition to supporting the N+M configuration, EMC Isilon supports an N+M:B option as well. As was stated earlier, M represents the number of drives or nodes that can fail and not impact data availability. Because the likelihood of a node failure is significantly lower than that of a drive failure, the :B option allows the administrator to override the node failure value of the protection level. So, in a 4+2:1 configuration, the system can tolerate two drive failures or 1 node failure. While this may not seem like a good option at first, it does provide a significant amount of capacity savings by reducing the amount of protection information that is stored on the system.
EMC Isilon supports protection levels of N+2:1 and N+3:1 in addition to the options previously discussed.

2. Where the EMC Isilon solution may not be a good fit

Applications that have high single-stream I/O requirements cannot take advantage of the EMC Isilon cluster’s distributed model, and will generally perform better on a traditional scale-up system, such as the VNX. Also, small, random I/O is not optimal for the EMC Isilon. Small random reads do not benefit from the aggressive pre-fetching capabilities of the EMC Isilon cluster, and small random writes do not utilize the write coalescing capabilities.

Transactional applications, such as Oracle, SAP, SQL, Exchange and SharePoint, have I/O profiles that are not a good fit – single stream or small, random reads and writes.