Re: NDB communication slow ?

[Mikael Ronström]

Hi,

A non-replicated system always have a faster response time, particularly for
write queries that execute the full commit protocol. Reads sounds a bit much though.

The difference looks a bit big to me even with this background knowledge.

One parameter that can help delay time possible is ndb_force_send (an option
in the MySQL Server).

You can also try setting your processes (particularly ndbd in higher priority or even
use real-time scheduling, this can have a great effect in some cases).

Apart from that there is not that much to do unless you want to try very advanced stuff on
OS scheduling and RT OS and so forth.

There is work ongoing to do more
at once which decreases response time in general (have seen 3-5 times improvements
in some queries in 5.0). You can try out SCI hardware if the problem is burning your feets
which is sure to decrease your response time by decreasing cost of communication.

In some cases there are also ways to write the query to decrease the number of turns to
take to the data nodes of the cluster.

Rgrds Mikael


Balasubramanian Azhagappan wrote:
> Hi everybody,
>
> we tested 2 cluster configurations:
>
> ndb2: Machine A (nodegroup 0)
> ndb3: Machine B (nodegroup 0)
> ndb4: Machine A (nodegroup 1)
> ndb5: Machine B (nodegroup 1)
>
> Both Machine A and Machine B has the same
> configuration.
> Each ndb nodes has 4GB RAM.
>
> 1. I made the nodes 2 and 4 down. With only 2 NDB
> nodes in the cluster (no redundancy). The answer
> time was 2-4 ms for each request (read or write).
>
> 2. Then again made nodes 2 and 4 UP. Then with 4
> NDB Nodes on 2 computers (with redundancy), we got
> 11-17 ms for each requests (read/write).
>
> I am puzzled of why i get such a variation in the
> request time. Is that cluster is taking much of
> the time in redundancy and data handling between
> nodes in different machines ?
> and How can we can reduce this? any suggestions
> please..
> Thanks in advance.
>
> Bala