carrium

Many people are confused by these numbers and that’s normal because of the tooling used, but let’s settle some basics: Let’s start with this. There is simply no way a Turbine-based shredstream can beat a multicast-based shredstream, and this is mainly because of two main reasons. 1.- Who performs the traffic replication: Multicast is not quicker than unicast at the transport level, don’t get confused by that. However, the key difference is who performs the replication of that point-to-multipoint data traffic that shred distribution follows. Today, this is done on the validator side by the OS network stack (CPU-based), which is slower than performing it at the ASIC level of the network switches in a multicast network. 2.- How Turbine fanout works: It is well known that Turbine uses a stake-based tiering system where shreds go first to the 200 most staked nodes, and then those nodes forward them to the rest, creating a few hops that can become a global ping-pong. If traffic replication is done at the network device layer, you don’t really need those tierings; the multicast network will spread this information at once. Some related factors: 1.- Transport efficiencies. Here you can benchmark the efficiencies of the public internet vs. private networks, or even vs. overlays over the public internet. In general, a well-meshed private network is more reliable due to reducing the amount of hops and events that occur on the public internet. Also, people forget that many hosting providers have private networks to connect their locations and remember for example that TeraSwitch runs 33% of the stake and they use their own global fabric to interconnect their locations. So if the nodes are colocated in these locations, you are not really comparing DoubleZero vs. the public internet, but also DoubleZero private network vs. TeraSwitch private network. Result in this specific situation? Very little difference. This is not applicable, of course, for all locations, which is why when you compare one option vs. the other, you really need to understand what you are comparing. 2.- Alpenglow: Solana will receive a major upgrade soon, and Rotor, the core protocol component in charge of shred distribution, will completely change how Turbine operates. In Rotor, the validators will delegate shred distribution to a set of relayers deployed by Anza. This should reduce the time required by validators to replicate the unicast traffic, and it should also make propagation quicker because the network will no longer be layered in stake tiers, so it will be more similar to multicast traffic flow than today with Turbine, but still over the internet. As a summary, the difference between Turbine based shredstreams vs Multicast based shreadstrams will be almost unfair to compare for the reasons shared above. Remember than today only 20% stake has been onboarded and we can perceive already big differences.