Hi ext3-gurus,
I was doing some testing of our 100 Mbit ethernet at work, and found a
weird problem which appears to be ext3 releated as far as I can tell.
The test is very simple. I use ncftp to connect to another machine
running vsftpd and download a very large file over the network. The
network is switched, and I get 8-10 MB/s transfers usually.
Now, if I download to /dev/null ("get -z bigfile /dev/null"), the
transfer is fast and steady. Same goes if I download to an ext2
partition. When I mount the partition as ext3 (with the default
data=ordered or data=writeback), ncftp stalls badly (progress display
stops updating for several seconds) and transfer rate drops to 4-5 MB/s.
vmstat traces show that the number of interrupts per second drop from
8-9K to a couple of hundred when ext3 begins writing out data.
I have tested on several different machines and configuration, the
pattern seems to be the same. The distribution is either RH 7.3 or RH
8.0, and I've tried both the latest errata (2.4.18-27) and the latest
RawHide (2.4.20-2.54) kernels.
While writing this report, I realized that I had only run my test on
machines with IDE partitions and 3Com NICs. For kicks, I tried on a
completely different machine with a SCSI disk and Intel NIC. Curiously,
this machine did not exhibit any stalls, even with ext3 in data=ordered
mode.
I've attached a longish file with vmstat traces of all the different
test cases, as well as selected dmesg output for the two test machines.
Any ideas?
Thanks,
Troels Walsted Hansen
Hi ext3-gurus,
I was doing some testing of our 100 Mbit ethernet at work, and found a
weird problem which appears to be ext3 releated as far as I can tell.
The test is very simple. I use ncftp to connect to another machine
running vsftpd and download a very large file over the network. The
network is switched, and I get 8-10 MB/s transfers usually.
Now, if I download to /dev/null ("get -z bigfile /dev/null"), the
transfer is fast and steady. Same goes if I download to an ext2
partition. When I mount the partition as ext3 (with the default
data=ordered or data=writeback), ncftp stalls badly (progress display
stops updating for several seconds) and transfer rate drops to 4-5 MB/s.
vmstat traces show that the number of interrupts per second drop from
8-9K to a couple of hundred when ext3 begins writing out data.
I have tested on several different machines and configuration, the
pattern seems to be the same. The distribution is either RH 7.3 or RH
8.0, and I've tried both the latest errata (2.4.18-27) and the latest
RawHide (2.4.20-2.54) kernels.
While writing this report, I realized that I had only run my test on
machines with IDE partitions and 3Com NICs. For kicks, I tried on a
completely different machine with a SCSI disk and Intel NIC. Curiously,
this machine did not exhibit any stalls, even with ext3 in data=ordered
mode.
I've attached a longish file with vmstat traces of all the different
test cases, as well as selected dmesg output for the two test machines.
Any ideas?
Thanks,
Troels Walsted Hansen