Read acceleration in a non-uniform cluster
Recently Luca Dell’Oca published an excellent article on how FVP behaves in a non-uniform cluster. This cluster contains hosts with flash resources and a host without a flash device. The article zooms in on write acceleration and it shows that once the virtual machine is migrated to the host without a flash resource, the write performance is degraded.
One of the features this article highlights is the automatic switch from write-back to write-through without any required input from Luca. FVP detected that there is no flash device and therefor it couldn’t accelerate writes. Immediately FVP switched the write policy to write through. Everything is automated and FVP immediately reconfigured the write policy to write back when it was moved to a host with a flash resource. For more information about this behavior, please read “Fault Tolerant Write Acceleration”.
One thing that Luca didn’t cover in-depth was the sustained read acceleration, even if the virtual machine is moved to the host without a flash device. It would have been very interesting if the hit rate metrics had been included, especially the network flash read. As FVP is aware of every version of every single block of data, a host running the FVP software can still retrieve the data from any host inside the flash cluster with or without flash resources present.
When the VM issues a read operation of data already present in the flash cluster, FVP will fetch the data from the previous source host and send it to the virtual machine. As there is no local flash resource present, it cannot store it locally, nevertheless FVP continues to fetch data from the flash cluster to keep the latency to a minimum while reducing the overall stress and load on the storage array.
Please read the following articles If you want to learn more about remote flash access;