Re: 64 node Oracle RAC Cluster (The reality of...)

On 06/21/2005 11:53:35 PM, Kevin Closson wrote:

> this is not true for a real CFS. A proxy-cfs or nfs
> exhibits the characteristics you fear, but not a
> fully symmetric, concurrent read:write CFS. Demand
> paging is nothing more than the internals of mmap
> which in turn is really nothing more than an IO.
> Oh, with one exception, it is entirely 100% read only
> (a major text fault that is). That concern is a
> red herring. Binaries execute just fine from a CFS.

Kevin, I heard your name before and I am fully aware that you know much more then me. Truly symmetric clustered FS that I used to work with was known as Files-11. DEC representative used to tell us, I quote, "not to do image activation from remote nodes". As you probably know, VMS did not do file caching before version 6. The only question I have is synchronization of caches across the nodes. How can you achieve speed similar to the local FS? OS utilities usually do not support direct I/O. Oracle released a plug-in replacement for Linux utilities, so that you have ftp, cp, dd, ls, tar and cpio being able to utilize direct I/O. The thing still doesn't work for scp and sftp. So, how do you get around synchronizing caches on different nodes? I assume that there is a concept similar to SCN which gets increased with each transaction and if local SCN is higher then the global one, then you know that there was a change on your side and send all buffers with the increased SCN to the other guy or write the blocks down to the disk and have the other nodes re-read them (OPS technique). It still looks like a significant overhead which would slow down normal file operations significantly and make things like "vi $ORACLE_HOME/network/admin/listener.ora" fairly expensive. That would also interfere with the database operation and compete for the same bandwidth, wouldn't it?

-- 
Mladen Gogala
Oracle DBA


--
http://www.freelists.org/webpage/oracle-l