Date: Thu, 5 Mar 1998 19:03:16 +0000 (GMT) From: Terry Lambert <tlambert@primenet.com> To: remy@synx.com (Remy NONNENMACHER) Cc: tlambert@primenet.com, hackers@FreeBSD.ORG Subject: Re: Disk write caches Message-ID: <199803051903.MAA29505@usr06.primenet.com> In-Reply-To: <Pine.A32.3.91.980305094557.16895O-100000@rs1> from "Remy NONNENMACHER" at Mar 5, 98 09:53:10 am
next in thread | previous in thread | raw e-mail | index | archive | help
> > > > I would be very interested to see what the benchmarks on a machine > > > > with soft updates would be with a disk with and without write > > > > cacheing. > > > > > > I tested cache benefits within disks. Write cache disabling lowered the > > > perfs from 2 to 10 time depending on operations. > > > > Did this include soft updates, however? > > No. It was using 3.0 from December (25). How does soft update compares > with async mounts? For many operations it's faster. Async mounts do not necessarily imply write gathering. Soft Updates use an I/O "clock" that can result in gathered writes. The difference is that there are a reduced number of I/O's in the softupdates uses of async I/O vs. the async mount uses of async I/O. For some operations, the clock and the number of buckets in the clock cause a slowdown. This can be tuned two ways (neither of which are implemented as tunables at this time). The reallocation is a bit of work, so it would require an event rather than a simple value change to make it work. The sysctl() code is not very well suited to this type of thing (attaching code to a change operation). In practice, these tunables would only be needed in pessimal cases, for benchmarking things like large tree creates (which are a tiny bit slower than async). If you combine the create/delete time, then the soft updates code is *vastly* faster than async, even with noatime, because it gathers metadata writes. > Will the driver immediatly responding 'OK' for each write have no > more interest with soft update ? I don't understand the question. The reason I asked my question (which Julian has stated he will be answering, thereby letting me off the hook of getting my own hardware that will disable write cache on the drive) is that I question the value of a write cache, if sync writes are not being performed for nearly any operations. Effectively, soft updates creates a write cache in software, which is smarter about OS usage patterns than the write cache in hardware (being that soft updates has an exact graphical model of the optimal usage pattern). My question bears on the utility of cacheing and returning a driver level write in the case where nearly all controller level writes are cached and returned by the soft update dependecy creation process. There has been a lot of interest in the Simon/SCSI/RAID thread about fault tolerance. It seems to me that a system *without* disk write caching is more fault tolerant than a system *with* disk write caching. It is a pity that standard PC architecture machines do not include power supplies with seperate AC and DC powerfail notification, nor motherboard circuitry to notify the OS of the AC failure so that the OS can do something about it (in a driver). Terry Lambert terry@lambert.org --- Any opinions in this posting are my own and not those of my present or previous employers. To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199803051903.MAA29505>