TruCluster Server Handbook (HP Technologies)

13.8 Direct Access Cached Reads

In V5.1A, when a cluster member performs a read from a file greater than or equal to 64KB and that file is on an AdvFS file system on a device physically connected to that member, then the read will be performed directly. Metadata operations and log updates will continue to go through the CFS server; writes will also go through the CFS server unless the file is also opened O_DIRECTIO. Figure 13-4 illustrates the I/O flow within V5.1A. By performing reads directly, traffic across the cluster interconnect can be reduced.

Figure 13-4: Direct Access Cached Reads

If the member does not have direct access to the device containing the AdvFS file system, reads will still be issued to the devices (and not to the CFS server) as if those devices were directly connected to the system as illustrated in Figure 13-5. It is the DRD subsystem that will send the read requests to a cluster member with direct connectivity to the storage devices containing the file system. In other words, the request will still be serviced remotely but at a lower level in the I/O subsystem hierarchy.

Figure 13-5: Direct Access Cached Reads – No Physical Connection

Direct Access Cached Reads are enabled by default, and currently there is no subsystem attribute available to modify this setting although applications that use Direct I/O will not use Direct Access Cached Reads.