Messages in this thread |  | | Date | Sat, 7 Dec 1996 00:35:52 -0800 (PST) | From | Michael Neuffer <> | Subject | Re: Linux decent performance |
| |
On 6 Dec 1996, Andi Kleen wrote: > > > > I don't know about asynchronous I/O and locking shared memory, but linux > > > > does support raw partition devices, and always has. > > > > > > Linux doesn't support raw devices that aren't cached in the buffer > > > cache. I think he means this. > > > > The buffer cache should only increase performace. Unless Sybase uses its > > own caching algorithm, in which case the buffer cache is redundant. Is > > this the case? > > The raw devices do lazy writing so that the DB can't be sure if the > data actually hit the disk. This lowers the reliability of the DB > in case of a crash.
Even worse, if you are trying/have to do share the physical storage devices, ie. multiinitator SCSI.
We are currently working on a highspeed distributed database engine.
We will have pairs of two hosts accessing two RAID arrays, one of which the host "owns" and beeing the backup for the other in case of the other host goes down.
The problem here is that you need to be sure that all relevant data is beeing written to the disk the moment and the way you want it. Otherwise you can not do a secure takeover in case of a failure or recovery.
On the other hand you obviously do not want to open a data file with O_SYNC for performance reasons. Only certain pieces of data need to be written synchronously the rest can be more or less savely be cached.
Even if you would do it, you would never know, when you access a drive if you get the data that you read really from the drive or from that the buffer/pagecache THINKS is on the disk, but which might be two entirely different things at that point of time, because the other host has updated the datastructure.....
Mike
Michael Neuffer i-Connect.Net, a Division of iConnect Corp. mike@i-Connect.Net 14355 SW Allen Blvd., Suite 140 503.641.8774 Beaverton, OR 97005
|  |