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
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)
21500
20900
2.3
Observed at 60 clients and 12 agents
40100
39300
Read operation (4k) ***
50100
46300
1.0
Read operation (20k) ****
5450
5400
3.1
Observed at 30 clients and 12 agents
10600
10100
3.2
Observed at 50 clients and 17 agents
15000
14700
2.9
** 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