s-tikhomirov
commented
4 months ago A few suggestions from my part.
Mesh figure
On the yellow envelope and arrows as edges:
- If P1 is the initial publisher, the arrow with the envelope shouldn't arrive into P1 from the left.
- If anything, I'd put the envelope near the (1,3) edge, however...
- ... if the envelope represents relaying the message, then it should be on (1, 2) edge as well.
- Should we reflect the fact that the message is also relayed along the route (1, 2, 6, 7, 8), it just arrived later? (Likely not for the sake of clarity, although better to mention it in the caption or the main text to avoid making a wrong impression that a message is forwarded via one path only, like a Lightning payment).
From the general principle of simplifying things:
- we have two times "D = 2": I would remove them altogether, as the caption says that this graph has D = 2
- (I'm more open for discussion on this one): I would also remove the "clouds" that represent P1's and P8's local meshes. It's unclear to me what idea we convey here - it's kinda clear from the edges themselves that P1 is connected to P2 and P3 (analogously to P8).
More generally, I worry that the "local mesh" term might confuse people. Yes, it's a term from the original GossipSub paper, but I'm not sure it's widely used elsewhere in this meaning. IMO, we can say: "mesh topology" (a network where all nodes have a similar degree); "mesh connection" (a type of GossipSub connection that behaves a certain way, an opposite of "gossip connection"), and maybe "the mesh" (all mesh connections - although I'd prefer "the mesh network" maybe). Instead of "P1 mesh" I would say "P1's mesh neighbors" or even "P1's neighbors" if it's clear from the context that we're talking about the mesh network. In the context of the figure in question, I'm not sure what's the value add of depicting the "P1 mesh" and "P8 mesh".
Indices: I suggest we change the notation for latencies using this principle: the lower index shows what happens (g - generation, v - validation, t - transmission), and the upper index shows where it happens (for generation and validation it is an index of a node, for transmission it's "A -> B" (or "A-B", or "A,B" - we can discuss), where A and B are the indices of the nodes, or abbreviations of cities).
A note on the generation latency: for uniformity, shall we add there an upper index for the publisher node (L_g^1).
Message hops figure
My suggestions would be:
- To swap "Bandwidth ampl." and "N1=1000..." legends. As they stand now, the "Bandwidth ampl." legend sticks visually with the dashed lines, which made me think for a moment that the dashed lines are amplification. Also, swapping the legends may avoid the upper-right legend being crossed by the amplification line. I'd try to avoid that crossing in any case.
- Can we spell "amplification" fully in the legend? Perhaps with a line break.
- (I'm less sure about this): shall we align the Y-axes? Currently they are nearly aligned (aligned means 4 on the left is opposite 4 on the right, etc), but not exactly. We should show clearly that the bandwidth line corresponds to the right Y-axis, and other lines correspond to the left Y-axis. If the two Y-axes are aligned, then it doesn't matter whether to look left or right if you want to know what is the value at a particular X-value. On the other hand, the two Y axes are for physically different metrics (an integer vs a coefficient)... that's why I'm not sure if aligning them is a good idea. Maybe, on the contrary, we should show them as unaligned for that reason.
- Bandwidth amplification should also be mentioned in the title?
Distribution figure
The main issue I see here is that there is a lot of things going on, and the reader might get lost. In particular, there are a lot of "variable = value" captions, which I think are better represented in other forms (such as a table).
My suggestions:
- remove "1) mu =..." lines from the captions. Instead, put these numbers into a separate table. Explain in the text what mu and P95 mean.
- remove "2) min=..." from the captions. AFAIU, the maximums duplicate the rightmost column in Table 3. (I don't understand tbh why the numbers (477, 945, 1689, 2468) are called max in the figure, but are just L_total in the table.)
- All that's left in the captions above the graphs would be: "Message size = X KB" ("KB in caps).
- remove "cumulative" from Y axis in the bottom row.
- remove repeated "1) single host 2) multi host" from within the graphs: we can describe this in the common caption: "The simulation (solid bars) and the multi-host measurement (dashed line) latencies..."
Sidenote: we shouldn't refer in the text to colors on the diagram because a) some people will print the paper in black-and-white, b) some people are color-blind. Ideally, all lines on the chart should be distinguishable only by solid / dashed lines, etc with color as an additional distinguishing factor.
Fix plots for paper