Has anyone actually seen performance degradation when a provider runs old silicon? Like measurable, not just "feels slower"?
Intel discontinuing my VPS CPU generation, should I worry?
Security patches are my concern. Intel stops microcode updates for EOL generations. Your provider can still apply OS-level mitigations but the really nasty stuff needs firmware.
HasanIce said:
Measurable
Yes, but not from age. From contention. An E5 v4 node with 2019-era density (4:1 oversubscribe) feels fine. The same node at 2024 density (12:1) because the provider stopped buying new hardware is where you feel it.
The CPU did not get slower. Your slice of it got smaller.
iBGP, eBGP, don't care, just peer
I moved from Contabo to Hetzner Cloud last year. The difference in compile times for Go binaries was roughly 40%. Same core count, newer architecture. CXL and faster memory helps.
steveipw
Member
Server33 said:
E5-2630 v4
That is a 2016 part. 10 cores, 20 threads, 85W TDP. For a VPS host they probably give you 1 vCPU which is one thread. You are sharing that 2.2GHz base clock with everyone else on the socket.
The chip is fine. Your neighbors are the variable.
zfs send | zfs receive. repeat.
This. I had a Contabo VPS that was great for 8 months then suddenly terrible. I bet they packed more tenants on the node.
Is there any way to check how many neighbors you have on a KVM host? I know you can see steal time in top but that only shows when they are active.
josemcelroy
Member
webxzhang said:
Check how many neighbors
Not directly. You can infer from /proc/cpuinfo siblings count versus what you paid for, and from timing attacks if you are bored and slightly evil. But no clean number.
Steal time over 5% consistently is your signal to move.
I just checked my Vultr instance. AMD EPYC 7713. Compile speed is ridiculous compared to my old Contabo box. Worth the extra $2/month for me.
What region? I am looking at Vultr Tokyo and they only list "High Performance" not the specific CPU.