Skip to content

Budget split: compute here, storage there, network elsewhere

VPS Hosting by armstrongvds 5 replies 202 views
#1

Been running a fun little distributed lab across three providers!

Compute layer: 2x Ampere A1 @ Vultr (Tokyo) — 4 cores, 24GB RAM total, insane value at $6/mo each. Runs circles around the x86 boxes I had before.

Object storage: Hetzner cold tier, €0.003/GB. Mounts via rclone.

Network/VPN front: RackNerd wireguard hub in EU, 10Gbps port.

Latency between layers is 180ms worst case but for my use case that's fine! Each piece does what it does best instead of one overpriced box trying to juggle everything.

Anyone else doing this? What combos are you running?

one small ping for man...
#2

Wah this one very complicated lah. Why not put everything in one server? Can save money also dong. Your setup already spend how much per month?

#3
armstrongvds said:
Each piece does what it does best

Bro this is we make a test with split architecture before but is very good for learning! I use Vultr too, their network is fast. How you sync the storage? rsync or what?

#4

Dig +short vultr.com
203.0.113.47

TTL on that A record is 300, not great for failover. Check your resolver — some ISPs cache longer than advertised. Propagation to my test resolver took 312 seconds, not 300. Your WireGuard endpoint will flap during any migration. Docs: https://www.wireguard.com

#5

Ngl this seems like massive overkill for anything that isn't training LLMs. What's the actual workload? Inference on those Ampere chips with 24GB shared is rough for anything over 7B params. VRAM math doesn't work out unless you're quantizing hard.

4 #6

Bro server down in tokyo last week since 3 hours, no? I see Vultr status page say Network Issue. Your Important data safe or not? Since you split everything, maybe storage still up. This why I like single provider, less headache.

Post a reply

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

Post reply Preview Save draft