NetBSD-Users archive
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]
Network very very slow... was iSCSI and jumbo frames
Hello,
Now jumbo frames are enabled :
23:27:55.227343 IP legendre.systella.fr.58386 >
euclide.systella.fr.iscsi-target: Flags [.], seq 3959352:3968300, ack
2593, win 4438, options [nop,nop,TS val 192 ecr 896198099], length 8948
23:27:55.227350 IP legendre.systella.fr.58386 >
euclide.systella.fr.iscsi-target: Flags [.], seq 3968300:3977248, ack
2593, win 4438, options [nop,nop,TS val 192 ecr 896198099], length 8948
23:27:55.227359 IP legendre.systella.fr.58386 >
euclide.systella.fr.iscsi-target: Flags [.], seq 3977248:3986196, ack
2593, win 4438, options [nop,nop,TS val 192 ecr 896198099], length 8948
But jumbo frames and Intel adapter don't increase NAS throughput :
204Mb 407Mb 611Mb 814Mb
0.99Gb
└─────────────┴───────────────┴───────────────┴───────────────┴───────────────
legendre.systella.fr => euclide.systella.fr 74.9Mb 78.2Mb 78.0Mb
<= 379Kb 410Kb 374Kb
With old Realtek and MTU 1500, throughput was about 10 Mbytes/s. With a
new Intel adapter and jumbo frames, I obtain the same result (screen
capture : https://hilbert.systella.fr/public/nas.png). Please not that
NAS CPU is not overloaded.
On NetBSD side, top returns :
load averages: 1.28, 1.24, 1.59; up 1+12:45:13 23:34:36
72 processes: 70 sleeping, 2 on CPU
...
Memory: 7767M Act, 3800M Inact, 16M Wired, 94M Exec, 10G File, 22M Free
Swap: 16G Total, 16G Free
PID USERNAME PRI NICE SIZE RES STATE TIME WCPU CPU COMMAND
21113 bacula-s 115 0 64M 6300K tstile/1 4:08 18.41% 18.41% bacula-sd
434 root 83 0 68M 5684K netio/6 4:51 5.32% 5.32% bacula-fd
0 root 0 0 0K 23M CPU/7 108:10 0.00% 0.68% [system]
1141 root 85 0 117M 2736K nfsd/0 89:02 0.00% 0.00% nfsd
1257 root 43 0 756M 522M parked/7 3:19 0.00% 0.00% istgt
467 root 95 -20 35M 4864K select/0 2:55 0.00% 0.00% openvpn
...
This system is not overloaded. NAS and NetBSD server are directly
connected (no switch). I have tried to add some options to ifconfig
(tso, csum...) without result.
I suppose there is somewhere a bottleneck, but where ?
JKB
Home |
Main Index |
Thread Index |
Old Index