FAST bandwidth savings and CPU

, ,
1

Imported from previous forum

2

Hi, I am about to embark on a series of tests to determine whether sending less data over network results in CPU reduction and how much. I think I can increase current bandwidth savings with FAST by 20% - 30% percent but the key requirement is also to bring down CPU utilization.
Has anyone performed such tests? For example, assuming the rate of 15K/messages per second, how is CPU utilization affected by the message size of 50 vs 100 bytes?

Thanks,
Dimitry

3

The main savings that you’re likely to see with smaller packets over Ethernet will be because of packet batching. If you can send more data in the same number of packets i.e. the same number of kernel calls, then you’ll get CPU reduction.

However, more messages per packet means higher latency for the first oldest messages, so it’s always a trade off.

4

The main savings that you’re likely to see with smaller packets over
Ethernet will be because of packet batching. If you can send more data
in the same number of packets i.e. the same number of kernel calls, then
you’ll get CPU reduction.

However, more messages per packet means higher latency for the first
oldest messages, so it’s always a trade off.

I plan to get reductions by utilizing across-message data compression. That is, using deltas I will send 1 byte instead of 20, for example. Thus, initial reduction will not introduce any additional latency and will use the same number of system calls as before. So the question is how much CPU if any sending 50-byte messages vs 1000-byte message at 15K messages/per second saves?

Thanks,
Dimitry