With reference to my earlier posting in this thread (Dec 16, 2002), where I quoted some timings for tar and dump on ext and reiserfs, I received a mail from Mr. Stelian Pop, the maintainer of BSD/dump, who noted that tar is actually cheating when outputting to the bit bucket. I've checked it up and found out that tar indeed skips a lot if the target ("-f") is either "/dev/null" or redirected to "/dev/null". My suspicion that there is something wrong with IPC in Linux was unwarranted. The reason why tar is so much slower when it writes through a dd pipe is that it then needs to output all the data without skipping anything. To better understand the issue look at this table of backup speeds (in MB/s) and ratios:
ext3 reiserfs ratio Backup method\From (fastdisk) (slowdisk) reiserfs/ext3 ------------------ ---------- ---------- ------------- tar > null 23.5 66.7 2.84 dump > null 19.4 -- ? dd > null 32.2 23.4 .73 tar | dd > null 10.3 8.8 .85 dump | dd > null 18.7 -- ? It shows that dump is a quite a bit more efficient backup tool than tar when tar doesn't cheat. A tar from ext3 is some 15% slower than a dump if we factor in the relative speeds of the two disks. That it is 2.8 times faster on reiserfs than on ext3 when it basically skips all data and only traverses the directories proves my point from previous postings - reiserfs is a quite a lot more responsive file system for meta data manipulations than ext3. Seeing how much handshake and massaging of meta data is caused by SMB I am prone to believe that reiserfs is ideal fs for samba. I hope to be able to fill in the blanks under dump/reiser entries in about a month and a half and then we shall review the data. My prediction is that a good reiserdump would be streaming at about 35 MB/s or better on the faster of the two IDE disks. -- To unsubscribe from this list go to the following URL and read the instructions: http://lists.samba.org/mailman/listinfo/samba