RE: load-testing with testlvs & ipvs 1.0.4

[Peter Mueller <pmueller@xxxxxxxxxxxx>]

> > director = 2.4.19-rc1 + [base redhat .config] + ipvs patch 
> + julian's hidden
> > patch + heartbeat + ldirectord.  2-700mhz P3 & 512mb of 
> ram, 7200rpm scsi
> > disk, etc.  (valinux 2230).  << I am using 2.4.19-rc1 
> because I have eepro
> > NICs >>.
> 
>       One of the most meaningful parameters: the PCI bus speed?
> 

I believe one is ethernet pro (the one I'm using) is onboard and the other
is standard 32-bit PCI slot.

cat /proc/pci
  Bus  0, device  11, function  0:
    Ethernet controller: Intel Corp. 82557/8/9 [Ethernet Pro 100] (rev 8).
      IRQ 18.
      Master Capable.  Latency=64.  Min Gnt=8.Max Lat=56.
      Non-prefetchable 32 bit memory at 0xf4200000 [0xf4200fff].
      I/O at 0x2800 [0x283f].
      Non-prefetchable 32 bit memory at 0xf4000000 [0xf40fffff].
  Bus  0, device  14, function  0:
    Ethernet controller: Intel Corp. 82557/8/9 [Ethernet Pro 100] (#2) (rev
8).
      IRQ 21.
      Master Capable.  Latency=64.  Min Gnt=8.Max Lat=56.
      Non-prefetchable 32 bit memory at 0xf4203000 [0xf4203fff].
      I/O at 0x2840 [0x287f].
      Non-prefetchable 32 bit memory at 0xf4100000 [0xf41fffff].

> > real server = 2.2.x, apache, 2 proc, 512mb, requests set to 
> reject according
> > to README file.  specifically, srcnet is changed to 
> 10.10.0.1 and then
> > make).  I issue the command "route add -net 10.10.0.0 
> netmask 255.255.0.0
> > reject" to drop/reject the packets from LVS-director during test.
> 
>       You can even use ipchains just to drop the flood. The
> route reject command wastes CPU in the RS.
> 

OK, I was just following the first example I saw.  I assume the cycles are
wasted because REJECT sends a packet back saying "no SYN for you!" (or
something like that), right?

>       May be you have PCI 33MHz, ipchains rules, etc. The other
> problem could be the routing cache performance. You can reduce the
> number of clients to 10 just to see what happens.

I think I have PCI 33mhz.  The director is 2.4.x kernel with essentially no
iptables rules.

Reducing the clients seemed to help all the machines become *MUCH* more
responsive..

<director>
 0  0  0      0 199904 105580  50512   0   0     0     0  103     7   0   0
100
 0  0  0      0 199900 105584  50512   0   0     0     5  105     9   0   0
100
 0  0  2      0 199900 105584  50512   0   0     0     0 22946  3478   0  49
51
 0  0  1      0 199900 105584  50512   0   0     0     5 41835  6295   0  91
9
 0  0  1      0 199900 105584  50512   0   0     0     5 41988  6349   0  91
9
 0  0  1      0 199900 105584  50512   0   0     0     0 41875  6350   0  92
8
 0  0  1      0 199900 105584  50512   0   0     0     7 44765  4417   0  83
17
 0  0  1      0 199900 105584  50512   0   0     0     0 42143  6668   0  86
14
 0  0  1      0 199900 105584  50512   0   0     0     5 42010  6660   0  87
13
 0  0  2      0 199900 105584  50512   0   0     0     5 42102  6651   0  87
13
 0  0  1      0 199680 105584  50512   0   0     0     0 42000  6359   0  87
13
<real server>
[root@stage-fe2 testlvs-0.1]# ./show_traffic.sh 
2 packets/sec
2 packets/sec
3 packets/sec
2 packets/sec
9 packets/sec
4945 packets/sec
18951 packets/sec
21056 packets/sec
21109 packets/sec
21057 packets/sec
21051 packets/sec
21045 packets/sec
21042 packets/sec
21037 packets/sec

Is this PCI bus the cause for 87% utilization at ~21000 packets/sec?

Thanks Julian & company.

Peter