usm bench

Quick health-check benchmark for a fresh VM or container. Not a serious benchmarking suite — this is a 30-second sanity test to catch "spot VM claims 64 cores but has IOPS like a USB stick" cases.

usm bench                     # quick mode (~30s)
usm bench --full              # bigger sizes, more samples
usm bench --no-net --no-gpu   # offline / non-GPU machines

What it measures

Section	What
CPU	Single-thread float Mops/s (tight inner loop in Python); core counts; load averages
Memory	Total / available / used; bytearray copy throughput for 256 MiB (quick) / 1 GiB (full)
Disk	Sequential write + read of 256 MiB (quick) / 1 GiB (full) to $TMPDIR. fsync before measuring write throughput.
Network	ping 1.1.1.1 min/avg/max; HTTPS download of a 10 MB Cloudflare speed test
GPU	If torch + CUDA available: matmul TFLOPS for fp32 and fp16 on a 4096² (quick) or 8192² (full) matrix

Skip flags

--no-cpu, --no-mem, --no-disk, --no-net, --no-gpu for any subset.

Honesty notes

• The CPU number is Python op rate, not native FLOPS. Comparable across machines for the same Python version; not comparable to other tools.
• Disk throughput depends heavily on $TMPDIR (tmpfs vs SSD vs spinning rust). Don't compare numbers across machines unless their $TMPDIR backends match.
• Network test goes to the public Internet — your VM's egress shaping, not local NIC speed.
• GPU FLOPS via torch.matmul is roughly representative of dense training workloads, not arbitrary kernels.

For real benchmarking: fio, iperf3, sysbench, mlperf, cublas-bench.

Source

scripts/bench.py