# DAS modeling and simulations

## Dissemination of rows and columns to validators

***Baseline assumptions (this is how the current model works, but almost none of these are cast in stone):***
* Block data is encoded using 2D RS coding from 256x256 to 512x512
* Dissemination is done using GossipSub channels per rows and per column
* Validators instruct their beacons nodes to subscribe to the channels corresponding to selected rows and columns
* Beacon nodes subscribe to the union of channels of all their validators
* Reconstruction is done at the beacon node (not validator) level
* Reconstruction is done by beacon nodes on-the-fly, injecting missing and reconstructed blocks in their respective channels when possible
* Attestations are done based on the information received this way, no further sampling is performed
* Validators are also listening when not attesting, and consider this in their later attestations

### Find baseline node to validator ratio (and distribution) to use in our simulations
* Goal: determine realistic and parametrizable distribution for number of validators per beacon node
* Methodology: reference data points from crawler -> analytical model
* Data
* Number of nodes in the network: 3500 ~ 5500 (Based on Migalabs)
* Number of nodes in the network: ~10K total ~7K Synced
* Number of validators: 450K ~ 500K (Based on beaconcha.in)
* Model
* Model different node-to-validator distributions as a superposition of validator "categories"
* 1st model:
* Staking pools: Gauss centered around 100?
* Solo stakers: Spike around 1 val/node?
* Ratio of the two categories

### Each validator selects X rows and columns based on local randomness (X=2 by default)

* Vary from 1 to 512? (Some unrealistic configurations)

[csaba] Q: With 5K-ish nodes and 400K-ish validators, and with 2\*512 channels, how many nodes will be on a given row (or column) channel?
Probability of node $n$ with $v$ validators participating in a given channel $C$:
$$
P(n \in C) = 1- \left(1-{X \over 512}\right)^v
$$

### What is the effect of X on attestations, bandwidth, and latency
* Metrics: what are the right metrics to measure attestation and networking performance?
* attestation:
* attestation ratio: ratio of attestations to validators in committee
* attestation delay distribution: sum<=1, with 1-sum being the ratio of no attestations
* attestation "precision": attestations should reflect the availability of the data. Similar to a statistical test, TP/TN/FP/FN characterisation might make sense. TBD

#### Start with Jupyter node model (Assumptions, parameters, and metrics)
* Start with a direct flat (unrealistic) network?
* Random distribution assumed
* GossipSub delay distribution
* Push to K-mesh network
* Pull (IHAVE/IWANT) for tail distribution
* Network failure model: How much was sent out in the network?
* Network failure model: Network partitions?

#### Simulation
* Latency and bandwidth assumptions
* Cloud nodes: High bandwidth and low latency
* Home nodes: Lower network resources
* GossipSub assumptions
* Overlay stable (first model)
* Sybil Attack
* Different channels for rows/column distribution
* Network Mesh Dynamics
* Impact of many channels

* Configurations
* Desired Lowest Common Denominator (a.k.a. Raspberry Pis)
* Desired network composition (a.k.a. what can actually work?)
* Realistic network distribution

**Confirm validator rotation and duties**

### Assuming failures, how many reconstruction steps does it take to recreate the full block?
* When you have enough data to reconstruct you do and send as if you received it
* Parameter: How much of the block was sent?
* Parameter: How many nodes do not get gossipSub data from the X channel

### How long does it take to rebuild the block in the worst case? How to attack the structure?

### What is the worst case from the attestation / erasure coding / latency perspective?

### How much network bandwidth takes per node?

### How much erasure coding CPU load does it take per node?

### What is the optimal number of GossipSub topics?

### Can we suffer network splits? What parameters should we use to avoid them?

### How frequently should we change the topics? (Hours, days, weeks)

### What is the effect of distributing only half (or partial) rows and columns?

## Failure model

### Node proposer with bandwidth/network/congestion issues

* Effects:
* Sending partial data

* Mitigation:
* Use RS to reconstruct the missing data on arrival and propagate full data
* If proposer sends by sample and randomize sample by sample then we get better fault dispersal across multiple rows

## Open questions
* GossipSub message is a single sample or an entire rows?
*