Hi Martin, apologies for the rather late reply. I've tried to answer your questions inline. > > We have an application that has 10 billion small (256-4096 byte) > files. We do 10 reads for every write. We read and write each file in > its entirety. Only 1% of the files are hot; i.e. being read and > written in the same hour. > > Some questions: > > * Can I have 500 clients all mounting the file system simultaneously? should not be a problem. * Will my reads be primarily out of memory? Or am I going to be > limited by spindles? If your files are re-read, you can gain some serious performance with io-cache. else they would be served off the server VFS block cache. How many bricks/ram will I need so that I'm > mostly reading from memory? Depends really on the work load, how many files are in common set etc. You can start trying from 1 :) Is the cache write through or will a > write require a disk access on the next read? Depends on the translators you use. Currently write-behind is the write cache translator, which cut short's the write system call but flushes through the data at the same time. avati
