On Tue, 2013-12-10 at 23:47 +0100, Henrik Goldman wrote: > Hello, > > Several pages including > http://linux-iscsi.org/wiki/Targetcli#Create_a_backstore seems to > suggest that you should use iblock as the backstore for production > setups. > > However from our tests we see that performance is significantly > decreased compared with buffered fileio. In our tests we found a 2-4x > difference. > > I understand obviously that buffered io is less safe than unbuffered > but it's not clear for me why performance is so much different. > > One of the greatest features of fileio is that you essentially use all > your system memory as cache... something which is not very > highlighted. > You just answered your own question. ;-) Buffered FILEIO means all WRITEs go into the filesystem buffer cache, and only an explicit SYNCHRONIZE_CACHE or WRITE FUA (Forced Unit Access) from the client will force flushing of said WRITE data from buffer cache to the persistent backend device. > Personally I don't understand why buffered fileio is not highlighted > as it's clearly giving the best performance. > Because if/when the target power goes out, you loose all WRITE data in the buffer cache for a buffered FILEIO device that has not been flushed to the backend device. Whoooops.. A proper production solution here is using bcache with a SSD / flash device with write-back mode with IBLOCK export, which puts all WRITEs onto the fast flash persistent storage, thus avoiding the potential for data-loss during power-loss as with the buffered FILEIO case. --nab -- To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
