|
@trippehh | |||||
|
Linux kernel networking is amazing. Seeing 5GB/s on HTTPS in the lab here using only 4 intel cores @ 2,5Ghz, serving 10Mbps video streams (again mostly IO). That is 10Gbps/core of AES put onto the wire.
|
||||||
|
||||||
|
Kelly Sommers
@kellabyte
|
25. tra 2018. |
|
||
|
Seph
@josephgentle
|
25. tra 2018. |
|
What?? How? Is that using dpdk or something? O_o
|
||
|
|
||
|
|
Kelly Sommers
@kellabyte
|
25. tra 2018. |
|
No way. The need for DPDK is highly overrated. On an untuned kernel no less.
|
||
|
|
||
|
|
Seph
@josephgentle
|
25. tra 2018. |
|
One machine? Thats 30x faster than highly tuned c/c++ frameworks:
techempower.com/benchmarks/#se…
Which is suspiciously fast. How is it different? Whats going on?
|
||
|
|
||
|
|
Kelly Sommers
@kellabyte
|
25. tra 2018. |
|
This is just IO benchmarking but also don’t forget if I recall the techempower servers aren’t that powerful.
Remember modern computers are very fast. 20 cores, each able to do billions of operations/second.
|
||
|
|
||
|
|
Seph
@josephgentle
|
25. tra 2018. |
|
You get good throughput from big packets. But my understanding is that there's a fair bit of overhead per-packet, so processing small requests caps out at ~x00k req/sec. Hence redis and memcache both at 150k-500k req/sec. 17m network req/sec out of the box sounds high for linux.
|
||
|
|
||
|
|
Kelly Sommers
@kellabyte
|
25. tra 2018. |
|
Naa, this is about half of what the box can do. iperf shows me the box can do 1.5M TX and 1.5M RX.
|
||
|
|
||