Hotrod client performance

[chavarao]

Hi,

I have two hot rod servers setup in a cluster. I am bench marking  performance. I have  a client w in hard loop  doing 100000 puts and gets.

For puts it took 135 seconds , hot rod client is running on same machine as one of the server which is specified in hotrod server list .

But  when I run client on different machine which is on same subnet its talking 260 seconds plus.

I am not sure why its taking that long and how to see where is the time spent?

Is any one has this issue?

Thanks

[jacob1111]

Hi Nag,

I am very interested to setup such a benchmark as well.

I was able to setup a hotrod local cluster, but could not

set up a remote one.

What did you do to set up hotrod remote cluster?

How did you start the server?

How did you start the client?

What XML files did you change?

I could not find any place with consistent description of this procedure.

Once I set it up I will run put and get benchmarks

to measure the replication performance.

Best regards,

Jacob Nikom

[chavarao]

Hi Galder,

Can you please let me know how to trace it where the performance is?

Thanks

Chava

[jacob1111]

Hi Nag,

I was able to setup Infinispan cluster for embedded and HotRod access modes.

Yes, the performance is different. It is also different if you ran the HotRod server

on remote or local machine.

I found about 50% difference in performance:

Here are the data (use fixed size fonts to see the fiormatting):

Two clients, two remote HotRod Infinispan servers, two nodes

+-----------------------+------------+-----------+-----------+

| Number of inserted    | Object     | Object    | Object    |

| and retrieved objects | size 0.1K  | size 1K   | size 10K  |

|                       |  (sec)     |   (sec)   |   (sec)   |

+-----------------------+------------+-----------+-----------+

|            1000       |     2.226  |    2.438  |    2.797  |

|           10000       |    12.457  |   12.662  |   14.999  |

|          100000       |    54.778  |   57.963  |   73.021  |

|         1000000       |   466.090  |  495.431  |  624.543  |

+-----------------------+------------+-----------+-----------+

Two local clients, one local HotRod Infinispan server, one node

+-----------------------+------------+-----------+-----------+

| Number of inserted    | Object     | Object    | Object    |

| and retrieved objects | size 0.1K  | size 1K   | size 10K  |

|                       |  (sec)     |   (sec)   |   (sec)   |

+-----------------------+------------+-----------+-----------+

|              1000     |    0.917   |    1.178  |    1.244  |

|             10000     |    4.811   |    4.786  |    5.711  |

|            100000     |   25.696   |   26.379  |   36.461  |

|           1000000     |  215.771   |  238.529  |  337.343  |

+-----------------------+------------+-----------+-----------+

[mircea.markus]

when the embedded mode test was run, were the nodes colocated?

[jacob1111]

Hi Mircea,

Thank you for your reply.

In case of two clients, two remote HotRod Infinispan servers and

two nodes those nodes were different computers. The replication

between remote HotRod servers happened through the network

(1GigE), NOT through the local interface 127.0.0.1

Here is the depiction of the architecture for two clients two local ispn servers and two nodes.

Here is the architecture for two clients two embed ispn servers and two nodes.

Here is the architecture for two clients one local ispn server and one node.

Here is the architecture for two clients one embed ispn servers and one node.

Here are two graphs comparing performance if the first two and the last two

architectures.

Best regards,

Jacob Nikom

[chavarao]

Hi Jacob,

I am actually thinking of using two hotrodservers and 2 or more Hot rod clients.

I am thinking about creating my own balanacing strategy and use replciation mode . My balancing strategey on HR client will go to first local server and if its not available  go to second server( which is different machine).

Hi Mircea,

Do you see any problem with it?

Thanks

Nag

[chavarao]

I havent tried yet but any one see a  problem ?

[jacob1111]

Hi Nag,

I am also thinking about similar architecture.

Here is its depiction:

I wonder how Infinispan configuration file is going to look like?

What is the access mode: is it going to be repliaction, distribution or invalidation?

Regards,

Jacob Nikom

[chavarao]

Jacob,

I was thinking using replication .

I am only thinking changing hotrod client so

That instead of using round robin it will

Use custom BL Strategy which will always

Connect to my preferred  node .

[jacob1111]

Don't you think that distribution mode wilk be better?

Best regards,

Jacob