More performance numbers

Anshul ran some more numbers to see the max throughput and latency that we have observed with CTP3. Here are some numbers to share.

Note that these are the max throughput numbers and latencies observed at those values. If you reduce the throughput, then latency goes down. For example, in the Read 2k, if the CPU are not maxed out , then you can get latencies of about 0.6 ms or so. 

Nodes

Throughput(max)

Throughput(avg)

Latency(in Milliseconds)(avg)

Comments

Read operation (2k)

1 node

30000

28600

1.5

Observed at 53 clients and 12 agents

Read operation (2k)

2 node

58500

55500

1.6

Observed at 100 clients and 17 agents

Read operation (2k)**

3 node

64000

59000

0.6

Observed at 150 clients and 17 agents

Read operation (4k)

1 node

21500

20900

2.3

Observed at 60 clients and 12 agents

Read operation (4k)

2 node

40100

39300

2.3

Observed at 100 clients and 17 agents

Read operation (4k) ***

3 node

50100

46300

1.0

Observed at 150 clients and 17 agents

Read operation (20k) ****

1 node

5450

5400

3.1

Observed at 30 clients and 12 agents

Read operation (20k) ****

2 node

10600

10100

3.2

Observed at 50 clients and 17 agents

Read operation (20k) ****

3 node

15000

14700

2.9

Observed at 50 clients and 17 agents

 

** With all 17 dual core clients ( hardware I have) the 3 quad core server’s CPU was not saturated L . The numbers are at around 65% CPU utilization.

*** With the 17 dual core clients the 3 node server CPU was not saturated. The numbers are at around 75% CPU utilization.

****In all 20K numbers the network was saturated.

Server Configuration : Intel(R) Core(TM) 2 Quad CPU Q9300 \8 GB RAM \ 64 bit win 2k3\1GBPS network.

Thanks

Murali