Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
Lars Ellenberg schrieb:
> uhm, what does your cpu say to dd if=/dev/hda of=/dev/null ?
> or to dd if=/dev/zero of=scratchfile ?
root at via1:/mnt/nb0# dd if=/dev/zero of=scratchfile
2488220+0 Records ein
2488219+0 Records aus
1273968128 bytes transferred in 156,543794 seconds (8138094 bytes/sec)
CPU on via1 85-95%
CPU on via2 30-40%
(identical hardware on both nodes)
Here the CPU consumption seems to be the limit, or maybe ext3. With drbd
full-sync I get 11.5 MB/sec -> 50% CPU on both nodes. Obviously drbd
full-sync is _very_ efficient.
Do it reverse:
root at via1:/mnt/nb0# dd if=scratchfile of=/dev/zero
2488220+0 Records ein
2488220+0 Records aus
1273968640 bytes transferred in 25,523378 seconds (49913794 bytes/sec)
CPU 70%
Here the disk is the limit, 50 MB is identical to hdparm.
> this is most likely NOT drbd consuming cpu, but _tcp_ used by drbd...
Yes, drbd0 consumes the half directly.
> and might be reduced considerably by using NICs that do the tcp work
> themselves (and nic drivers that are aware of it).
Don't know.
# lspci -v
00:14.0 Ethernet controller: Advanced Micro Devices [AMD] 79c970
[PCnet32 LANCE] (rev 36)
Subsystem: Allied Telesyn International AT-2450 10/100 Fast
Ethernet
Flags: bus master, medium devsel, latency 32, IRQ 10
I/O ports at e000 [size=32]
Memory at df002000 (32-bit, non-prefetchable) [size=32]
Expansion ROM at <unassigned> [disabled] [size=1M]
Capabilities: [40] Power Management version 1
For my purposes the configuration is faster than needed. The hosted low
traffic websites, 1 GB traffic per month, produce significantly below
10% CPU.
> (and probably by using zero copy network writes as we do in 0.7)
Benchmark figures for the audience?
Helmut Wollmersdorfer