More nodes spend more time

classic Classic list List threaded Threaded
6 messages Options
Level D Level D
Reply | Threaded
Open this post in threaded view
|

More nodes spend more time

Hi,

I demand both read and write operations, so I chose partitioned mode cache. 

When I increased my cluster by one node, I got slower reads.

Is there any way to make it faster?


vdpyatkov vdpyatkov
Reply | Threaded
Open this post in threaded view
|

Re: More nodes spend more time

Hello,

Need more detail: which task you try to solve, how you measure delay on read and so on.

Obviously you can get networking overhead for data which stores on remote node. 
But compared to REPLECATED cache you do not waste time copying data to all nodes on write.

Can you provide example?

On Wed, Jun 22, 2016 at 10:23 AM, Level D <[hidden email]> wrote:
Hi,

I demand both read and write operations, so I chose partitioned mode cache. 

When I increased my cluster by one node, I got slower reads.

Is there any way to make it faster?





--
Vladislav Pyatkov
Level D Level D
Reply | Threaded
Open this post in threaded view
|

Re: More nodes spend more time

In reply to this post by Level D
Thanks for asking.

I'm trying to get faster reads. Write is good for now.

Here's my read example. It's just testing environment.

I have two caches, Cache1 and Cache2. And preload 45,000 tips in Cache1. Put 1,000,000 tips into Cache2 to trigger continuous query in Cache1. After the end, console will print the total time.

When I increased my cluster from two nodes to three nodes, console returned me a larger number.

Is there something more you need to know?
vdpyatkov vdpyatkov
Reply | Threaded
Open this post in threaded view
|

Re: More nodes spend more time

Hi,

Are you sure, you correct measure?
JVM need a warm-up before the code will otimizirovan.

What cluster topology: 
Where is fisical machines (witn Ignite node) located?
Where is located client node (where you do benchmark)? 

On Wed, Jun 22, 2016 at 11:38 AM, Level D <[hidden email]> wrote:
Thanks for asking.

I'm trying to get faster reads. Write is good for now.

Here's my read example. It's just testing environment.

I have two caches, Cache1 and Cache2. And preload 45,000 tips in Cache1. Put 1,000,000 tips into Cache2 to trigger continuous query in Cache1. After the end, console will print the total time.

When I increased my cluster from two nodes to three nodes, console returned me a larger number.

Is there something more you need to know?



--
Vladislav Pyatkov
Level D Level D
Reply | Threaded
Open this post in threaded view
|

Re: More nodes spend more time

In reply to this post by Level D
I'm quite sure.

All machines are in the same computer room.

I run my only one client code on one of my servers.

In addition, I use ignite 1.6 and execute sql query in filter.

What does "JVM need a warm-up before the code will otimizirovan" exactly mean?
What shoulde I do about the warm-up?

Er...

Question is not that complicated.It's not about continuous query.
Can you explain why more nodes spend more time when I execute sql query in partationed mode?
Theoretically more nodes should be faster?

vdpyatkov vdpyatkov
Reply | Threaded
Open this post in threaded view
|

Re: More nodes spend more time

Hi,

Please, can you explain are comparing?
In first time you run client and server in same machine.
In second time you run client and one server in same machine and one server in another.
Is it correct?

If that true. In first time client communicate with server through shared memory (it quite faster then honest socket connection).
In second part of data go across network, that leads to increase overhead.

warm-up
I mean JVM must go through some iteration, before will show the best result.

It would help if you have sent the code showing the performance drop.

On Thu, Jun 23, 2016 at 4:31 AM, Level D <[hidden email]> wrote:
I'm quite sure.

All machines are in the same computer room.

I run my only one client code on one of my servers.

In addition, I use ignite 1.6 and execute sql query in filter.

What does "JVM need a warm-up before the code will otimizirovan" exactly mean?
What shoulde I do about the warm-up?

Er...

Question is not that complicated.It's not about continuous query.
Can you explain why more nodes spend more time when I execute sql query in partationed mode?
Theoretically more nodes should be faster?




--
Vladislav Pyatkov