My dude I pay more than that for bodega sandwiches
You want NVMe performance at bodega sandwich prices you gonna get bodega sandwich reliability
Not defending OVHcloud just saying expectations need calibration
My dude I pay more than that for bodega sandwiches
You want NVMe performance at bodega sandwich prices you gonna get bodega sandwich reliability
Not defending OVHcloud just saying expectations need calibration
Ok sale you also on OVHcloud now? I am curious — your sustained 60MB/s vs pablowild 80MB/s, is different tier or different luck?
I am think maybe the throttle is not fixed number, is "fair share" dynamic based on neighbor noise. So your 60MB/s is because someone else is hammering same array.
This is worse than fixed limit. Fixed limit you can plan around. Dynamic "fair share" is just random number generator.
I benchmark Vultr block storage again last night. Their "NVMe" block storage is actually consistent 3000 IOPS, no cliff. More expensive yes but at least is real number.
I am tell you now what they say, save you time: "The resources advertised correspond to the hardware specifications of the platform. Actual performance may vary based on workload and shared environment factors."
I am have this exact ticket with HostHatch 2022. Same copy-paste, different logo.
The only way to know true disk is bare metal or KVM with passthrough. LXC is always mystery meat.
This. So much this.
I moved my stuff to KVM at Hetzner specifically because of this. Their CX line is still "shared vCPU" but at least the disk is qcow2 on NVMe with no funny cgroup throttles that I can find. Sustained performance matches burst.
More expensive than OVHcloud? Sure. But I know what I'm buying. With OVHcloud and these LXC-style VPS plans you're buying a lottery ticket that says "maybe NVMe, maybe not, maybe 400MB/s, maybe 40MB/s, good luck."
Random number generator with extra steps. And no seed value you can control.
Following this thread because I am consider OVHcloud for GPU instance storage. But if their block storage is this "fair share" nonsense, then local SSD only.
Does anyone know if their "Local NVMe" on higher tier is also throttled? Or is that actually real?
I have their Advance-2 dedicated in London. The local NVMe there is genuine — no cgroup throttles, sustained 2000+ MB/s sequential. But that's dedicated, not VPS.
For their Public Cloud "Local NVMe" flavors I cannot say. Different platform. The VPS line we're discussing here is definitely the bottom tier with the mystery meat disk.
My advice: if you need predictable storage for GPU workloads, bare metal or at least dedicated instance. Shared VPS disk is always going to be "best effort" no matter what marketing calls it.