Added EKS Infrastructure

[utf94]

1. EKS infrastructure is added.
2. Dockerise the entire application
3. Added build configuration
4. Added Load balancer

[utf94]

The PR is still in progress. Different GPU machines will be tested and the best results will be shared with @avaer and on approval, those build commands will be added.

Current buildspec points towards personal AWS of Kaneki.ken which will be updated once the machine spec is finalised

[utf94]

https://preview.webaverse.com/?url=https://webaverse.github.io/assets/male.vrm&ext=vrm&type=png&width=200px&height=100px

Render time 2 seconds.

Horizontal scaling integrated that spins upto 7 pods in 0 additional seconds.

https://preview.webaverse.com/?url=https://webaverse.github.io/assets/male.vrm&ext=vrm&type=png&width=200px&height=100px

[avaer]

$ time curl 'https://preview.webaverse.com/?url=https://webaverse.github.io/assets/male.vrm&ext=vrm&type=png&width=200px&height=100px'
Warning: Binary output can mess up your terminal. Use "--output -" to tell
Warning: curl to output it to your terminal anyway, or consider "--output
Warning: <FILE>" to save to a file.

real    0m4.234s
user    0m0.000s
sys     0m0.016s

[avaer]

The current time is about 5 seconds to render. Is there a comparison table of the performance of each instance type?

[avaer]

Any way to not hardcode the account id?

[utf94]

Yes, we will move these files to Cloud Formation and the account id will be dynamic then.

[avaer]

The code looks fine.

[avaer]

I think we can merge this, but I'd like to understand the different GPU options, and see if we can get the render time cut in half to about 2s as stated.

[utf94]

@avaer Actually the time which is getting taken is in the downloading of the asset. If we host the asset files in S3 or in a CDN we will get rid of 1.5 seconds easily. Right now the GPU consumption is very less. In order to further check and refine it. I will add logs throughout the preview and screenshot system to know which part takes what time and also will play with the flags of the puppeteer if all go in vain then we can test more GPUs as testing GPU comes with a high cost since AWS has limited GPU to P3.2Xlarge machines (at minimum)

[ahadshams]

@rob-webaverse can you also confirm the impact on infra cost after this?

[rob-webaverse]

@rob-webaverse can you also confirm the impact on infra cost after this?

It's adding about $4200/month right now, so more than +100% previous costs.

[avaer]

If we host the asset files in S3 or in a CDN we will get rid of 1.5 seconds easily.

How? The first hit is the common case and it seems uploading to CDN would make it even slower by adding that layer.

Also I don't fully understand the timing. This file is 8KB and fits in 6 packets. I doubt 6 packets takes1.5 seconds, unless AWS routing is really that slow.

[avaer]

Btw I think optimizing this is pretty important for the user experience. Seconds shaved off will make for a much better gamefeel.

[utf94]

@avaer we are having a render time of 1.2 seconds. While the asset download time is 2.1 seconds what are your suggestions on this?

That makes the request span between 4->5 seconds in total.

[avaer]

Which part of the download is slow? That seems excessive.