Edit this page

Technique

23 DEC 2017

Q: If the worker gone and hub has spare slots, will the hub automatically reassign tasks? Will the client notice it?

A: In MVP will. The problem is that we're building a general purpose computer, which can potentially run any docker containers. Some tasks are cannot be continued in case of failure, some tasks can, but they need to transfer volumes to the new container. So it's not a simple question, whether we could implement this option. Still we can. We can do some thing similar to Docker swarm or Kubernetes services. Or even implement Kubernetes on top of SONM network. In future releases"


Q: In the future, the ultimate solution is build a overlay network between hub and workers, in that way neigher upnp, port forwarding or public IP is required, hope will available in official release August 2018.

A: Exactly. It's one of the crucial features we gonna implement in spring. Besides, overlay network will also be accessible from client's side. So imagine you draw a 3D scene in 3DMax and want to render it using RedShift or Octane. You install local render farm manager, such as Deadline with SONM plugin, and just click a button in 3DMax to start rendering. This render farm manager can buy instances in AWS - and so it can do it with SONM - and he will buy tens, hundreds or event thousands of rigs automatically and run redshift in docker containers. Each container will be linked to an overlay network and will be accessible from render manager, so it can split the task into pieces and run them simultaneously on all rigs bought. That's one of use cases in detail we gonna implement.


Q: What is the purpose of IPVS?

A: The purpose of IPVS is to allow, for example, web services, running on workers without public IP. Actually, if container hosted on worker doesn't want to expose public ports and thus doesn't need public IP address, there is no need of IPVS or other solutions. So in future versions we can try to configure routers to proper ports translation, or we can switch off IPVS, and do no not allow workers without public IP to run tasks that expose ports.


Q: Can you a possibility of making hubs redundant and load balanced?

A: Yes, we can. Consul wth Raft is used to make fail over on Hub.


Q: How customers specify multiple hubs' ip addresses in workers' configs and about clients' side.

A: It's another way. SONM uses internal discovery service. Each hub announces its IP or IPs to this service, associated with its EthID. When Client or Worker want to connect to Hub, they send discovery service Hub EthID and receive its IPs to make connection.


Q: If the hub goes down and there is another associated with EthID, will the worker automatically switch to it so the client won't notice any failures?

A: Yes, it will. If the connection is dropped, worker will try all Hub's addresses to reconnect. Still tasks will continue running on workers.


Q: Won't UPnP enabled in router be enough for worker?

A: For production release in August - surely. For MVP we need to check. I have to clarify. Ordinar miner, say, rig owner, will have to start both Hub and Worker components of SONM. Worker shall be started on every PC you want to share resources. This PC can be behind the NAT - no problem. Hub is one, or better three PCs, that will work as a supervisor for your workers. Consider it to be a small mining pool. So these PCs should have white IPs. May be, we can configure router to passthrough some ports, but this is to be checked.


Q: What does it mean that SONM is an alternative for mining?

A: The problem is that running computing tasks for money on your computer is not actually mining. it's like mining - because your earn money, but it's not about searching hashes, checking blockchains, etc. So we are renaming miners to workers to precisely reflect what they are doing.


Q: What are minimum requirements for sonm?

A: If you want to mine, you need almost any amd64 Linux box, even virtual. But customers will want to run different task with certain requirements. We'd like to test SONM on GPU-intensive cases such as rendering, machine learning. So PC with Linux and 2-3-... Nvidia Geforce 960-1080 cards, i5 CPU, 16G RAM, 200G HDD is quite good.


Wait a new publication of Q&A!