Re: Benchmark results questions

[Ray <ra...@cisco.com>]

Hi Dmitry,

It's been a while now, did you check out what happens?




--
Sent from: http://apache-ignite-users.70518.x6.nabble.com/

Re: Benchmark results questions

[dkarachentsev <dk...@gridgain.com>]

Hi Ray,

I've finally got results of query benchmarks:
4s1c 80725.80 80725.80
4s2c 78797.90 157595.80
4s4c 54029.70 216118.80

8s1c 64185.60 64185.60
8s2c 61058.10 122116.20
8s4c 34792.70 139170.80

First column - cluster configuration (in 8 server variant 2 nodes per
machine), second - average throughput per driver, third - total throughput.

So there is some degradation on larger cluster. First of all, queries does
not scale, you need broadcast the same request to all nodes. Adding more
nodes doesn't make this job easier, but even harder, because client has to
process more requests and responses - that's what I see from profiler. In
larger cluster client consumes more CPU, more IO and increases pressure to
GC.

On the 8s4c configuration two servers on single machine influence on work of
each other. Opposite to SQL, single put/get requests are targeted on primary
node only, not to all, and adding more nodes gives almost linear increasing
of throughput.

You will get increasing of query performance on large grid with much bigger
dataset and more complex SQL. In that case processing would be spreaded
among servers and data digesting will be parallelized.

Thanks!
-Dmitry



--
Sent from: http://apache-ignite-users.70518.x6.nabble.com/

Re: Benchmark results questions

[Ray <ra...@cisco.com>]

Hi Dmitry,

Thanks for you results.
I'll try to increase the number of clients as well to see if the throughput
can increase.

Can you also test the sql query throughput?
In my result, it seems with more servers the sql query throughput goes down
dramatically.
Is there a valid explanation for this phenomenon?

Thanks



--
Sent from: http://apache-ignite-users.70518.x6.nabble.com/

Re: Benchmark results questions

[dkarachentsev <dk...@gridgain.com>]

Hi Ray,

I wasn't able to run benchmarks quickly, but I've got following results for
atomic put throughput (numbers are a bit lower that could be, because of
profiling):

Throughput	        Cluster
264930	 	1c4s
513775		2c4s
968475		4c4s

281425		1c8s
530597		2c8s
905491		4c8s

In 8 servers configuration used two nodes on single machine. Results say
that cluster is scaling normally, but clients unable to load it fully. It's
could be seen from CPU usage: with 1 client 7%, with 4 clients - 30%.
Also Yardstick shows results per one driver. 

Try to run each server node and driver on a separate machine, and increase
number of drivers.

Attached results for 4 server configurations with your settings 
atomic_put.zip
<http://apache-ignite-users.70518.x6.nabble.com/file/t552/atomic_put.zip>  

Thanks!
-Dmitry



--
Sent from: http://apache-ignite-users.70518.x6.nabble.com/