Hello, a few comments on your tests: - Reading from /dev/urandom could be a bottleneck, depending on how that random data is generated. Best to avoid this, if you need random data, try to use a bench tool that can quickly generate dynamic random data.
- Writing to ZFS can give all sorts of results, i.e. it may be doing compression, encryption, deduplication., etc. You'd need to disable all those features in order to have comparable results to NetBSD local file system. - I think by default, dd does not call fsync() when it closes its output file, with GNU dd you need to use conv=fsync argument, otherwise you could be benchmarking writing data to OS page cache, instead of virtual disk. On Tue, Mar 20, 2018 at 9:20 AM, Chavdar Ivanov <ci4...@gmail.com> wrote: > Well, testing with a file of zeroes is not a very good benchmark - see the > result for OmniOS/CE below: > ---- > ➜ xci dd if=/dev/zero of=out bs=1000000 count=1000 > 1000+0 records in > 1000+0 records out > 1000000000 bytes transferred in 0.685792 secs (1458168149 bytes/sec) > ---- > > So I decided to switch to previously created random contents and move it > with dd between two different disks. Here is what I get: > --- > --------Centos 7.4 -------- XFS > ➜ xci dd if=/dev/urandom of=rand.out bs=1000000 count=1000 > 1000+0 records in > 1000+0 records out > 1000000000 bytes (1.0 GB) copied, 9.6948 s, 103 MB/s > ➜ xci dd if=rand.out of=/data/rand.out bs=1000000 > 1000+0 records in > 1000+0 records out > 1000000000 bytes (1.0 GB) copied, 2.49195 s, 401 MB/s > --------OmniOS CE --------- ZFS > ➜ xci dd if=/dev/urandom of=rand.out bs=1000000 count=1000 > 1000+0 records in > 1000+0 records out > 1000000000 bytes transferred in 16.982885 secs (58882812 bytes/sec) > ➜ xci dd if=/dev/urandom if=rand.out of=/data/testme/rand.out bs=1000000 > 1000+0 records in > 1000+0 records out > 1000000000 bytes transferred in 21.341659 secs (46856713 bytes/sec) > --------NetBSD-current amd64 8.99.12 ------- FFS > ➜ sysbuild dd if=/dev/urandom of=rand.out bs=1000000 count=1000 > 1000+0 records in > 1000+0 records out > 1000000000 bytes transferred in 32.992 secs (30310378 bytes/sec) > ➜ sysbuild dd if=rand.out of=/usr/pkgsrc/rand.out bs=1000000 > 1000+0 records in > 1000+0 records out > 1000000000 bytes transferred in 23.535 secs (42489908 bytes/sec) > ---- > > OmniOS/ZFS and NetBSD/FFS results are comparable, the Centos/XFS one is a > bit hard to explain. > > This is on the same Windows 10 host as before. > > Chavdar > > On Mon, 19 Mar 2018 at 23:16 Chavdar Ivanov <ci4...@gmail.com> wrote: > >> I ran my tests with our dd and also with /usr/pkg/gnu/bin/dd, supposedly >> the same or similar enough to the one in Centos; there was no significant >> difference between the two. The fastest figure came on the system disk when >> it was attached to an IDE controller with ICH6 chipset. about 180MB/sec. >> All other combinations return between 110 and 160 MB/sec. Tried >> with/without host os cache, also there is a setting that the disk is solid >> state. No apparent difference. >> >> My host system is build 17120, so that may explain something. Not the >> difference in figures though, comparing to Centos. >> >> Chavdar >> >> On Mon, 19 Mar 2018 at 23:06 <m...@netbsd.org> wrote: >> >>> On Mon, Mar 19, 2018 at 02:58:06PM +0100, Fekete Zolt?n wrote: >>> > Any setting which influence the test and I didn't apply? >>> >>> yes, need to figure out what to make GNU dd behave the same. >>> It has different defaults. >>> >>
