On 2020-11-16 2:53 a.m., Anders Magnusson wrote:
Yep, there is something that makes userspace extremely slow.Notice though that the kernel is actually faster with I/O than old kernels (like Ultrix). Me and Johnny tested that some years ago.
On the KA650 system it seemed Ultrix 4.5 has faster reads from the RD53 on RQDX3 and DEQNA.
For the RD53 I was using dd if=/dev/rra0c bs=128k count=100 of=/dev/null. All the NetBSD versions I tried read the RD53 at just over 300 kbytes/sec, and Ultrix was around 500 kbytes/sec. The Micropolis 1325 drive spec says 0.625 MB/S, though there's also seek time and rotational latency to consider.
Ultrix 4.5 could cat a 1 meg file from NFS to /dev/null in 2.6 seconds. NetBSD 2.0 took 3.5 seconds and later versions got progressively slower, with 9.1 taking over 8 seconds. I was hoping to use a binary search like git bisect to find one NetBSD version which made the DEQNA slower, but the change was gradual over many versions. Though interestingly only later versions do caching which makes subsequent reads of the same file fast.
I still wonder if something is wrong with the DEQNA driver. There were no more of those "qe0: xmit logic died, resetting" errors, but I notice that the 9.1 boot has little pauses while loading the kernel. Occasionally it pauses so long it times out or otherwise fails. I've never seen Ultrix netload or Athena etftp do that, and I don't think earlier versions of NetBSD boot did that.