10 messages in com.mysql.lists.clusterRe: SCI Transporter with 4.1.12
FromSent OnAttachments
Kenji HIROHAMA19 May 2005 06:27 
Mikael Ronström20 May 2005 03:02 
Kenji HIROHAMA22 May 2005 19:16 
Mikael Ronström23 May 2005 01:36 
Kenji HIROHAMA23 May 2005 01:47 
Mikael Ronström23 May 2005 03:48 
Kenji HIROHAMA23 May 2005 05:04 
Kenji HIROHAMA26 May 2005 06:57 
Mikael Ronström27 May 2005 12:44 
Matt Parzych28 May 2005 10:24 
Subject:Re: SCI Transporter with 4.1.12
From:Kenji HIROHAMA (kenj@gmail.com)
Date:05/23/2005 05:04:26 AM
List:com.mysql.lists.cluster

Hi Mikael,

OK, perhaps we should use SCI Sockets in our environment.

After buying dual CPU boxes, we will test SCI Transporters.

Thanks for your assistance!

Regards, Kenji

On 5/23/05, Mikael Ronström <mik@mysql.com> wrote:

Hi,

2005-05-23 kl. 10.48 skrev Kenji HIROHAMA:

Hi Mikael,

SCI Transporter is only for communication between Data nodes. We will use SCI Socket for interconnect between SQL and Data node in our environment.

ok

I understand that SCI Transporter is used only with dual CPU, and one of them CPU never stop working for Data node.

Still so, do you think, with single CPU, SCI Transporter cause worse performance even for interconnect between Data nodes?

Yes, most OS don't really like processes that run all the time so they can be scheduled away in a manner such that it does nothing for some time every now and then. Even with dual CPU this can sometime happen and the best is if it possible to lock the process to the CPU.

However, I haven't really made any tests on single CPU boxes, all of our tests were made on dual CPU boxes.

Rgrds Mikael

On 5/23/05, Mikael Ronstrom <mik@mysql.com> wrote:

Hi,

2005-05-23 kl. 04.17 skrev Kenji HIROHAMA:

Hi Mikael,

We would like to use SCI Transporter for interconnect, not SCI Sockets, simply for better performance. Our environment is; 4 single CPU servers for Data nodes 2 single CPU servers for SQL nodws Total number of servers is six.

With single CPU servers it is NOT recommended with SCI Transporters. A minimum for use of SCI Transporters is dual CPU's where one of the CPU's can be dedicated for the Data Node. For SQL nodes it is NEVER recommended to use the SCI Transporter. Thus it is only recommended with SCI Transporter in a narrow set-up for Data Nodes. Generally SCI Sockets is the recommended set-up.

Rgrds Mikael

According to the manual, you recommended to use dual CPU "dedicated" server for Data node with SCI transport. Should we use SCI Socket under single server?

Anyway, when compiling 4.1.12 source tarball, we had two problems.

1.)Using BUILD/compile-pentium-max As reported as a bug, we changed "check-cpu" part in BUILD/SETUP.sh, so we succeeded in compiling itself. However, MySQL Cluster didn't work with SCI dynamic libraries.

2.)Usign configure fire Even adding "--with-ndbcluster --with-ndb-sci=[path to SCI dir]", SCI dynamic libraries are not linked.

Before 4.1.11, there was no problem. There woudl be some problems in 4.1.12 source tarball, so I would like to ask who compilled successfully.

On 5/20/05, Mikael Ronstrom <mik@mysql.com> wrote:

Hi Kenji, You can always use the SCI Socket variant with the normal max binary if this brings problem. Can you describe what happens and possibly we could assist in finding out the problem.

Rgrds Mikael

2005-05-19 kl. 15.28 skrev Kenji HIROHAMA:

Is ther anybody who suceeeded in compiling 4.1.12 with "--with-ndb-sci=<path>" to enable SCI Transport Interconnect?

Even if we hard-coded "check-cpu" type to skip a bug reported here; http://bugs.mysql.com/bug.php?id=10696 , and compiled without any errors, SCISI libraries seems not to be contained.

Before 4.1.11, we didn't have any troubles.

Any helps?

Mikael Ronstrom, Senior Software Architect MySQL AB, www.mysql.com

Jumpstart your cluster: http://www.mysql.com/consulting/packaged/cluster.html

Mikael Ronstrom, Senior Software Architect MySQL AB, www.mysql.com

Jumpstart your cluster: http://www.mysql.com/consulting/packaged/cluster.html

Jumpstart your cluster: http://www.mysql.com/consulting/packaged/cluster.html