|
Hi Julian:
This is a proxy server. So the actual number of masqueraded
connections will be larger than the inbound connnections
reported by lvs.
I have seen this under actual client load, although I have
been inducing it recently with either a server near the cluster
(100Mps local net) or from a server remotely (dual T1 on the
remote system - cluster on a 10Mpbs feed). This system has been
live during these tests, so there is a consistent client load
of 120 active connections. The load I'm adding is (at a minimum)
200 simultaneous connections for 10000 connections.
-d
PS - Did you see F1 qualfying results this afternoon? Just curious.
-----Original Message-----
From: Julian Anastasov
To: Dan
Cc: 'lvs-users@xxxxxxxxxxxxxxxxxxxxxx'
Sent: 5/6/00 8:05 AM
Subject: Re: masquerading table slowdown
Hello,
On Fri, 5 May 2000, Dan wrote:
> Hi:
>
> My table size is 2^18. I have approximately 40M of free memory. When
my
> total number of active & inactive connnections (as reported by
ipvsadm)
> approaches 4000 the following message starts to appear in the log and
> connection speed slows dramatically.
Is this a FTP service? I can recommend you to change the MASQ
table size too (IP_MASQ_TAB_SIZE) but this is not related to the problem
with the kernel message you see. It is only for faster table lookups.
>
> kernel: IP_MASQ:ip_masq_new(proto=TCP): could not get free masq entry
> (free=36214)
>
> The system is running in NAT mode. Other parameters:
>
> LVS machine: PIII/500/64M with 4 Intel Etherexpress 10/100s (1
external IP,
> 2 private IP, 1 heartbeat). There real servers are PIII/500/256M with
2
> 3Com905 10/100s. There are 7 real servers. Both NICS on each real
server
> have traffic routed to them (evenly weighted). Only one of the private
NIC
> cards on the virtual server is used at any time (the other is a
failover).
Is the LVS attacked from real clients or from your test
suite near the cluster?
Regards
--
Julian Anastasov <uli@xxxxxxxxxxxxxxxxxxxxxx>
|
