storm-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Rui Abreu <rui.ab...@gmail.com>
Subject Re: Nodes underutilized
Date Wed, 07 Oct 2020 15:15:53 GMT
>
> 2) What might cause my high capacity (typically over 1) to not be
> offloaded to a more idle machine?


Are you consuming from Kafka?

But the machine which hosts the spout is pegged with significant load.


This hints at using more workers to spread the load and increasing the
spout's parallelism. But it depends on how many Kafka partitions your input
topic has.

On Tue, 15 Sep 2020 at 18:39, Thomas L. Redman <tomredman@mchsi.com> wrote:

> 2.2.0, I just upgraded not long ago.
>
> On Sep 14, 2020, at 9:28 AM, Rui Abreu <rui.abreu@gmail.com> wrote:
>
> Hi Thomas,
>
> Which version of Storm are you using?
>
> On Sun, 13 Sep 2020 at 20:23, Thomas L. Redman <tomredman@mchsi.com>
> wrote:
>
>> Sorry, I had previously sent this from a different email address, not
>> sure how well that would work with this service, hence this re-send.
>>
>> I’m running storm on a 3 node cluster, 32 physical cores in each node. I
>> have a complex topology with one spout which is a singleton, connected to
>> several other bolts most all of which doing natural language processing.
>> Most of these are pretty heavy weight. The input spout is easily capable of
>> outpacing the downstream bolts. I get good performance, but on only one
>> node, even though I specify 3 worker nodes for my topology. StormUI
>> indicates for any given component that the executors for that token on the
>> idle machines have emitted very few tokens, and have transferred none!
>>
>> When I look at the machine usage with htop, I see indeed only one of the
>> nodes is really getting any usage at all. My heaviest computation nodes
>> have a very high capacity value. But the machine which hosts the spout is
>> pegged with significant load. I have used almost exclusively shuffle(I
>> prefer localOrShuffleGrouping) grouping, but that doesn’t help. I will have
>> machines that are simply receiving few tuples to operate on, and those few
>> tuples are not transferred (and I admit I don’t know quite what that means).
>>
>> So, I have two questions:
>> 1) Why would a component on a node remote from the spout have a lower
>> Emitted count, and have a Transferred count always at zero?
>>
>> 2) What might cause my high capacity (typically over 1) to not be
>> offloaded to a more idle machine?
>
>
>

Mime
View raw message