Podman with CPU limit vs. KVM

[jb-alvarado]

Hello,

I did some tests with podman nginx container vs. a KVM VM with 2 threads. For testing I have two root servers and on one I run:

wrk -t12 -c400 -d30s http://...

On the other runs KVM and podman on debian 12 (podman version 4.3.1). Container is build with:

podman run --uidmap=0:33:256 --gidmap=0:33:256 --name=nginx -d \
  -v /opt/pods/data/nginx/www/html:/var/www/html/:ro --net vswitch \
  --network-alias nginx --ip=10.10.0.5 --hostname nginx \
  --cpus 2.0 -m 15g \
  nginx:latest 

Then I only change the --cpus value.

My results looks like:

VM 2 Cores

Running 30s test @ http://kvm.example.org/test.html
  12 threads and 400 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     6.03ms    2.24ms 218.32ms   85.25%
    Req/Sec     5.51k     0.91k   12.81k    68.89%
  1978941 requests in 30.08s, 439.72MB read
Requests/sec:  65797.99
Transfer/sec:     14.62MB

Container without limit

Running 30s test @ http://podman.example.org:8888/test.html
  12 threads and 400 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     2.98ms    1.42ms  85.09ms   88.21%
    Req/Sec    11.45k     1.51k   37.81k    72.90%
  4108632 requests in 30.08s, 0.89GB read
Requests/sec: 136580.49
Transfer/sec:     30.35MB

Container CPUS 2

Running 30s test @ http://podman.example.org:8888/test.html
  12 threads and 400 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    24.71ms   26.98ms 315.96ms   80.75%
    Req/Sec     2.03k     1.34k    6.64k    65.24%
  729457 requests in 30.09s, 162.09MB read
Requests/sec:  24239.87
Transfer/sec:      5.39MB

Container CPUS 4

Running 30s test @ http://podman.example.org:8888/test.html
  12 threads and 400 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     9.19ms   11.51ms 275.57ms   86.61%
    Req/Sec     5.73k     2.82k   38.29k    62.60%
  2053390 requests in 30.10s, 456.27MB read
Requests/sec:  68225.65
Transfer/sec:     15.16MB

How come it looks like that KVM can better utilize the CPU and has also better latency? Do I miss something, or I understand the --cpus parameter wrong?