perf: linear SCTP reassembly, one SACK every second packet, 4 MB buffer in the CLI

aiortc walked a message's fragments after each one arrived (n^2/2 steps
for a 1 MiB chunk) and acknowledged every packet. Measured on loopback,
400 MB to the CLI: 19 MB/s before, 39 after (49 by the node's count);
uploads from the CLI 25 -> 29 MB/s.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 18 insertions, 0 deletions

diff --git a/packages/meshbay-cli/src/meshbay_cli/client/udp_buffer.py b/packages/meshbay-cli/src/meshbay_cli/client/udp_buffer.py
new file mode 100644
index 0000000..c0045d8
--- /dev/null
+++ b/packages/meshbay-cli/src/meshbay_cli/client/udp_buffer.py
@@ -0,0 +1,18 @@
+"""A larger receive buffer on the client's ICE sockets, applied on import.
+
+In a download the client is the receiver, and it reads its socket from Python
+— on the same event loop as the terminal interface — more slowly than a node
+sends in bursts. aioice asks the kernel for 256 KB per socket; what does not fit
+is dropped by this machine's kernel (`UdpRcvbufErrors`), the node's SCTP sender
+takes each drop for congestion and halves its window. Measured on loopback, 400
+MB: 3245 drops and 19 MB/s with 256 KB. The node does the same for the same
+reason in the other direction (`meshbay_node/transport/webrtc/udp_buffer.py`).
+
+Linux grants no more than `net.core.rmem_max`; Windows grants what is asked.
+"""
+
+import aioice.turn
+
+WANTED = 4 * 1024 * 1024
+
+aioice.turn.UDP_SOCKET_BUFFER_SIZE = WANTED