keelOS next to KVM
keelOS development build 5de4f7e (FreeBSD 14.5, bhyve), measured 3 October 2026, vs Proxmox VE 9.2 (Linux 7.0.2, QEMU 11.0, KVM), measured 1 October 2026
CPU
| Test | Unit | keelOS | KVM | keelOS / KVM |
|---|---|---|---|---|
| 7-Zip benchmark | MIPS (higher is better) | 23,267 | 24,533 | 95% |
| sysbench CPU | events/s (higher is better) | 4,769 | 4,905 | 97% |
| Build Redis (make -j4) | s (lower is better) | 42.61 | 39.96 | 94% |
| Process wake-ups (perf sched pipe) | s (lower is better) | 4.84 | 3.579 | 74% |
Process wake-ups. Waking a process on another vCPU costs more under keelOS's hypervisor; it shows in perf sched pipe and a little in the compile and compression tests.
Memory
| Test | Unit | keelOS | KVM | keelOS / KVM |
|---|---|---|---|---|
| sysbench memory | MiB/s (higher is better) | 5,272 | 5,182 | 102% |
Storage
| Test | Unit | keelOS | KVM | keelOS / KVM |
|---|---|---|---|---|
| 4K random read, QD32 | IOPS (higher is better) | 81,229 | 79,368 | 102% |
| 4K random write, QD32 | IOPS (higher is better) | 28,105 | 36,366 | 77% |
| 4K random read, QD1 | IOPS (higher is better) | 6,517 | 6,456 | 101% |
| 4K random read, QD1, p99 latency | us (lower is better) | 216 | 264 | 122% |
| 1M sequential read, QD8 | MB/s (higher is better) | 1,458 | 1,970 | 74% |
| 1M sequential write, QD8 | MB/s (higher is better) | 1,326 | 1,906 | 70% |
Random reads are on par with KVM since 3 October (QD32: 54% → 102%, QD1: 101%). Two things changed. bhyve's virtio disk used to cost one exit from the VM for every request; now a disk thread that has nothing to do watches the VM's request queue for a moment instead, as QEMU's iothread does. And each virtual disk has 32 worker threads instead of 8.
Random writes: 35% → 77% with one ZFS setting. ZFS sizes its dirty-data limit from the memory of the system it runs in. On keelOS that is the host's own share, not the whole machine (4 GB of 16 here), so the limit was 372 MB against 1.6 GB on the Linux host, and ZFS wrote a transaction group every 30 ms: too small to be efficient. With the limit at 1.6 GB, as in this run, the same test gives 77% of KVM; keelOS now sets it that way by itself (from the next build). This does not need more memory: the dirty data lives in the ZFS cache, which stayed within its ceiling throughout. The rest of the write gap, and the large sequential reads and writes, are still open.
A clock that ran fast. While measuring we found that bhyve's ACPI timer ran 0.07% fast. Linux VMs set their clock's speed by it when they start, later see that it does not agree with their other timer and fall back to a slow clock source; from then on every program that reads the time is slower (the same QD32 read test lost a third). Our run of 1 October was on that slow clock source. It is fixed in this build.
Network
| Test | Unit | keelOS | KVM | keelOS / KVM |
|---|---|---|---|---|
| VM → another machine (1 GbE port) | Mbit/s (higher is better) | 941 | 941 | 100% |
| Another machine → VM (1 GbE port) | Mbit/s (higher is better) | 933 | 941 | 99% |
| VM → VM on the same host, 1 stream | Mbit/s (higher is better) | 24,925 | 31,070 | 80% |
| VM ← VM on the same host, 1 stream | Mbit/s (higher is better) | 23,877 | 30,519 | 78% |
| VM → VM on the same host, 4 streams | Mbit/s (higher is better) | 25,934 | 26,991 | 96% |
| VM ← VM on the same host, 4 streams | Mbit/s (higher is better) | 23,482 | 27,018 | 87% |
Between VMs on the same host. KVM moves packets in the kernel (vhost-net) and passes large segments between VMs. keelOS's virtual switch is in the kernel too and passes the same large segments: 78–96% of KVM. What is left is in bhyve's virtio network card, which runs in a user-space process that every packet still passes; a path that stays in the kernel is planned. The physical 1 GbE port is full on both.
How the VM-to-VM rows were measured
On 2 October 2026, on the same server with the same two VMs, with keelOS's virtual switch: a switch in the kernel (netmap's VALE) that moves packets in batches and passes 64 KB segments between VMs. We added VLANs to it and fixed three bugs in FreeBSD (in bhyve, in the network driver framework and in the ixl driver) that stopped it under load or on a physical port. It ships in 1.0; this run used a test kernel with the switch set up by hand, both VMs on access ports of one VLAN. Through the physical 1 GbE port a VM reaches the port's limit as before. On another boot of the same setup we measured 41–84% of KVM. On 3 October we found and fixed one cause of such differences between boots, a Linux VM falling back to a slow clock source (see Storage); these rows will be measured again. Before the switch, with tap interfaces on a bridge, the same test gave 24–26% (the last two columns).
| Test (Mbit/s) | KVM | keelOS | keelOS / KVM | before: taps on a bridge | before / KVM |
|---|---|---|---|---|---|
| VM → VM on the same host, 1 stream | 31,070 | 24,925 | 80% | 7,579 | 24% |
| VM ← VM on the same host, 1 stream | 30,519 | 23,877 | 78% | 7,473 | 24% |
| VM → VM on the same host, 4 streams | 26,991 | 25,934 | 96% | 6,837 | 25% |
| VM ← VM on the same host, 4 streams | 27,018 | 23,482 | 87% | 6,983 | 26% |
GPU passthrough next to bare metal
A GeForce GTX 1080 Ti passed through to a Debian 13 VM on keelOS, then the same Debian disk booted directly on the same server. Everything the GPU does by itself runs at full speed; copies between main memory and the GPU are about a quarter slower in the VM.
| Test | Unit | keelOS VM | Bare metal | VM / bare metal |
|---|---|---|---|---|
| GPU memory bandwidth | GB/s (higher is better) | 349 | 349 | 100% |
| FP32 compute | GFLOPS (higher is better) | 13,229 | 13,030 | 102% |
| FP64 compute | GFLOPS (higher is better) | 425 | 419 | 101% |
| INT8 compute | GIOPS (higher is better) | 11,522 | 11,027 | 104% |
| Kernel launch latency | µs (lower is better) | 3.92 | 3.82 | 97% |
| hashcat SHA-256 | GH/s (higher is better) | 4.13 | 4.089 | 101% |
| hashcat WPA | kH/s (higher is better) | 559 | 557 | 100% |
| Copy host → GPU | GB/s (higher is better) | 6.15 | 8.1 | 76% |
| Copy GPU → host | GB/s (higher is better) | 7.78 | 10.89 | 71% |
Supermicro X10DRU-i+, 2 × Xeon E5-2667 v3, 32 GB. Debian 13 (kernel 6.12), NVIDIA driver 550.163, OpenCL (clpeak) and hashcat's built-in benchmark. VM: 4 vCPU and 8 GB on the GPU's NUMA node; bare metal pinned to the same node. Median of three runs.
On bare metal the same copies drop to about 6 GB/s when the program runs on the other CPU, so keelOS puts a VM with a passthrough GPU on the GPU's NUMA node by itself. The copy gap is on our list for after 1.0.
Setup
- Server: Supermicro X11SPW, Xeon Gold 5122 (4 cores, 8 threads), 16 GB, two Samsung PM1725a NVMe drives. Why this one and not the server behind the numbers on the home page (2 × Xeon Gold 6144): the comparison needs the disks wiped and Proxmox VE installed, and this is the one of our two servers we can wipe; the other one runs our long tests.
- VM: Debian 13, 4 vCPU, 4 GiB, virtio-blk and virtio-net. The same disk image on the same ZFS volumes (volblocksize 16K, no compression), the same MAC address, the whole test disk written once before.
- keelOS: development build 5de4f7e (FreeBSD 14.5, bhyve), measured on 3 October 2026. Proxmox VE 9.2: kernel 7.0, QEMU 11.0, ZFS 2.4;
--cpu host, diskscache=none,aio=io_uring, iothread; vhost-net; measured on 1 October 2026, the server has run keelOS since. The same ZFS cache limit on both; no other VM running. - Storage on one NVMe drive on both sides (Linux could not read the second drive's protection information); the test disk written through once before the tests; the ZFS dirty-data limit at 1.6 GB on both (Linux's default for this machine; on keelOS set by hand for this run, automatic from the next build).
- Each test three times; the table shows the median and all three runs are in the raw data.
What we are doing
The virtual switch is being built into keelOS's network settings for 1.0. On disks, the rest of the random-write gap and large sequential I/O are next. Every release is measured again on the same server, and this page shows the latest run.