Spread requests over all available buckets.

[BleuSquid]

Smaller requests are quicker and allows requesting timed-out cells more quickly.

When no timed-out cells:

• No change on total network time
• Average 30% reduction on initial load as objects begin rendering as soon as the first request returns

With timed-out cells:

• Time to first re-request is greatly reduced.

[PrinterElf]

Is this increasing the number of requests?

[BleuSquid]

Only in scenarios where we are at a zoom level that requests between and 5 and 125 tiles.
If the request is for 36 tiles, we go from 2 requests (25, 11) to 5 requests (8, 8, 8, 8, 4).
If the request is for 125 tiles or more, there is no change because this already uses all the buckets. 5 requests @125 tiles

[jonatkins]

the real question is, how does it compare to the number of tiles per request/number of requests made by the stock intel map. that's what we should consider as acceptable behaviour, and when it was written it was a good match to stock intel. not sure if stock has changed since then though.

[BleuSquid]

I have analyzed the network behaviour of the current version: it already does not match stock.

Stock will only use as many connections as the initial request. I haven't logged and verified tile ids, but I believe each batch of tiles is locked to that specific request thread, and any time-outs are re-requested within the same thread.
IITC uses more connections on the subsequent request, I believe as a side effect of "if the tiles for this request have had several retries, use smaller requests"