garykwh
Member
Inference on a Budget
- Joined:
- Jun 2024
- Posts:
- 173
- From:
- San Jose, US
Ngl I just use Vultr's DNS (https://www.vultr.com). For LLM inference workloads DNS is like... the least interesting problem. My nodes need VRAM, not zone transfers. 8x A100s @ 80GB each, now thats worth optimizing
Provider DNS is fine until it isn't, then you learn BGP and anycast real fast I guess.
CUDA cores are my love language
ronwit
Member
Budget King
- Joined:
- Jun 2024
- Posts:
- 122
- From:
- Osaka, Japan
My nodes need VRAM, not zone transfers.
This. I'm running a Minecraft server on a $3 VPS in Osaka. I use whatever the provider gives me. My players care about TPS, not DNSSEC.
Though I did switch to Cloudflare's 1.1.1.1 for recursion because my ISP's resolver was caching stale records for like a day. (https://www.cloudflare.com)
instant noodles, instant deploys
pam3
Member
Latency Witch
- Joined:
- Aug 2024
- Posts:
- 74
- From:
- Melbourne, Australia
Recursion I outsource to Quad9
Same here. Malware blocking + no logging + anycast from Melbourne under 15ms. Hard to beat that running your own resolver unless you really want to.
I did try running unbound for a month. The cache hit rate was worse and I had to maintain the root hints. Why?
100% packet loss at hop 3
garykwh
Member
Inference on a Budget
- Joined:
- Jun 2024
- Posts:
- 173
- From:
- San Jose, US
Worry about your backups first.
Real talk. I see people in my GPU cluster Discord spending days on DNS infrastructure for a side project that has no users. Opportunity cost is a thing.
That said, I do run a local unbound on my inference nodes. Docker pulls a lot of domains, caching saves seconds on cold starts.
CUDA cores are my love language