Skip to content
Register

Compiling the Linux kernel across 4 cheap architectures

VPS Hosting by nerdbann 24 replies 2.9K views
nerdbann

Member OP

DNS Propagator

#11
nina64 said:
Your sample size is one compile per architecture

Fair cop. I re-ran ARM64 three times after reading this: 847s, 851s, 849s. Discarding nothing, average 849s. The TTL joke stands.

X86_64 on new node: 892s, 887s, 895s. Average 891s.

ARM64 still wins, but margin is 5% not 29%. Architecture gap is real, oversubscription gap was fiction.

irarose

Member

#12
playersofia said:
NAT is theft of address space

Theft requires a victim. RFC 1918 space was allocated for exactly this purpose. Your 2001:0db8::/32 is documentation space per RFC 3849, not routable. Even your examples are fictional.

My Contabo in Nuremberg gets native IPv6. My other Contabo in Nuremberg does not. Same account, different policy. This is the actual compromise.

sander_ams

Member

#13
nerdbann said:
X86_64 on new node: 892s, 887s, 895s

Same thing happened to me on same oversold neighbor issue on x86_vz, they moved me to St. Louis from New York and it got worse. Now on Sydney and stable. Contabo's Nuremberg and Sydney seem okay, US nodes are lottery.

svendeal

Member

Deal Sniper

#14

Which kernel config did you use

Still waiting on this. Defconfig? Tinyconfig? If you used 

make localmodconfig
the ARM64 advantage disappears because module count drops and compile time becomes I/O bound, not CPU bound. Contabo's ARM64 nodes — https://contabo.com/en/vps/ — are in Singapore or Sydney for APAC, x86_64 in same DC?

world record: 4min Arch install
engzur

Member

BGP Dreamer

#15
nerdbann said:
X86_64 on new node: 892s, 887s, 895s

Your traceroute changed. New node is on premium AS now, same as ARM64. I told you. They don't move you to "another x86 node," they move you to another tier's infrastructure when you complain loud enough. Check your ASN path.

playersofia

Member

#16
irarose said:
Your 2001:0db8::/32 is documentation space

Symbolic address. You knew what I meant. My actual prefix is 2a0e:dc0::/29 from Vivacom. Contabo gives me /64, I subnet it, I breathe. You in Nuremberg get native, you are lucky. Bulgaria gets what Germany throws downstream. This is hierarchy, not dignity.

RA

Member

4 #17

Contabo support told me they don't have "premium AS" and "budget AS." One network, different upstream blends per datacenter. My ticket was about CPU steal, not routing. They moved me within same DC, St. Louis to St. Louis.

SP

Member

#18
nerdbann said:
The TTL joke stands.

Now 300s gap, same hop count. My point stands. Latency is not your compile time. CPU is your compile time. Network people see packets and think everything is packets.

lee_mcr

Member

#19
svendeal said:
Which kernel config did you use Still waiting on this.

Contabo's 400GB VPS is what, 8 EUR? That's 24 EUR/year, 16p/GB/year. Your threshold is meaningless here.

I run kernel compiles on Kimsufi KS-1 for comparison. Atom N2800, 6.12 defconfig, 4h17m. Not in this benchmark because not Contabo, but someone asked about cheap architectures and nobody mentioned dedicated Atom.

Honey badger don't care... about downtime
nina64

Member

6 #20
rahul_astra said:
Same DC, St. Louis to St. Louis

Then engzur's AS topology theory is falsified. Single DC, single upstream blend, different hypervisor host. This is CPU scheduling, not BGP.

I want to see 

perf stat -e cycles,instructions,cache-misses make -j$(nproc)
from both nodes. Contabo uses KVM everywhere, but kernel version on host matters for scheduler hints.

Post a reply

You need an account to reply. Log in or register to join the conversation.

Post reply Preview Save draft