Re: Slow connect times across certain dbserveraliases
Posted in 1999
On Sat, 27 Feb 1999 07:22:06 +0000, David Williams
<djw@smooth1.demon.co.uk> wrote:
>In article <36d48a27.16849418@itsinet1>, Brad Coates <bcoates@cobe.com>
>writes
>>Has anyone ever seen the initial connection to a database take longer
>>using one dbserveralias as opposed to another on the same NIC? Here's
>
> Why more than one dbserveralias on one NIC? Multiple aliases are
> designed to be used to enable parallel access to >1 NIC. With >1
> alias per NIC seems like a bad idea although I can't think why,
> it just feels that you will get more overhead.
>
>
There's no network bottleneck at this point. We did try putting the
"offending alias" on a different NIC. It made no difference.
>>
> !!! I would expect it is due to different NETTYPE settings.
> Try applying the NETTYPE settings for rimstcp5 to rimstcp2 and
> see if things change.
All aliases are using the same NETTYPE settings.
> ?? Number of sockets per port seems as if you know how to delve
> into the kernel. What about number of networks packets in the
> in/out queues per port?
>
Did that. Found more packets going across "offending" port. The
catch is, once you're in, it's FAST. Seems like the poll threads can
handle the throughput ONCE you get in. I think a linked list is
getting too long and the connection process is scanning through each
slot trying to find where the next connection can be "inserted" for
processing.
>>Tonight, we're going to add more poll threads so the config will look
>>like the following.
>>
>>NETTYPE soctcp,12,150,NET>>
> It feels NETTYPE related. But more than one thread per physical CPU
> means that they all cannot be utilised at the same time..
>
Our NETTYPE has been changed to
NETTYPE ipcshm,1,50,CPU
NETTYPE soctcp,7,250,NET
That's 8 poll threads for 8 CPU VPs. Didn't make a difference.
I even wrote a script to randomly pick an alias to balance the load.
Didn't help. It just spread the problem.
We're going to try "Multiplexed Connections" to see if we can beat
this issue. I tried turning it on once but I had too many
dbserveraliases defined (I had 7 + 1 DBSERVERNAME). We can get to 10
supposedly in 7.31.
We may try something like Tuxedo to manage the connections since we're
unable to find a resolution. We're also going to purchase a V-Class
HP server. Nothing like throwing a bunch of hardware at the problem.
:)
Thanks for the help.