Brad Fitzpatrick
51682d245b
Each client connection ran two goroutines for its lifetime: the reader in sclient.run and a sendLoop blocked in a select over its send queues, pong, peer gone, mesh update, and keepalive channels. Almost all clients are idle at any moment, so the second goroutine mostly pinned memory: a 4 KiB stack, a g struct, a sudog per select case, three channels, and a context and errgroup. At 100k idle connections that was about 8 KB of a client's 22 KB RSS. Instead, start up the sendLoop only as needed, letting the goroutine go away otherwise, like Go 1.28-dev's http2 code (golang/go@5c51011e82) with similar parking to https://go.dev/cl/834084 but DERP's producers are all non-blocking, so a kick bit replaces that http2 code's send count. Measured with 100k idle TLS connections, server RSS per client went from 22.3 KB to 14.5 KB (22.9 KB to 15.3 KB after each connection had carried a packet), goroutines dropped from 2 to 1, and with no change in BenchmarkSendRecv throughput or allocations and no change in the time to do 100k serial round trips, each of which parks and wakes the writer. (it's super cheap to start goroutines) Updates #21064 Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com> Change-Id: I7c3e9a51d4b8f2607a1e5c3d9f8b2a4e6c0d1f3b
DERP
This directory (and subdirectories) contain the DERP code. The server itself is
in ../cmd/derper.
DERP is a packet relay system (client and servers) where peers are addressed using WireGuard public keys instead of IP addresses.
It relays two types of packets:
-
"Disco" discovery messages (see
../disco) as the a side channel during NAT traversal. -
Encrypted WireGuard packets as the fallback of last resort when UDP is blocked or NAT traversal fails.
DERP Map
Each client receives a "DERP Map" from the coordination server describing the DERP servers the client should try to use.
The client picks its home "DERP home" based on latency. This is done to keep costs low by avoid using cloud load balancers (pricey) or anycast, which would necessarily require server-side routing between DERP regions.
Clients pick their DERP home and report it to the coordination server which shares it to all the peers in the tailnet. When a peer wants to send a packet and it doesn't already have a WireGuard session open, it sends disco messages (some direct, and some over DERP), trying to do the NAT traversal. The client will make connections to multiple DERP regions as needed. Only the DERP home region connection needs to be alive forever.
DERP Regions
Tailscale runs 1 or more DERP nodes (instances of cmd/derper) in various
geographic regions to make sure users have low latency to their DERP home.
Regions generally have multiple nodes per region "meshed" (routing to each
other) together for redundancy: it allows for cloud failures or upgrades without
kicking users out to a higher latency region. Instead, clients will reconnect to
the next node in the region. Each node in the region is required to be meshed
with every other node in the region and forward packets to the other nodes in
the region. Packets are forwarded only one hop within the region. There is no
routing between regions. The assumption is that the mesh TCP connections are
over a VPC that's very fast, low latency, and not charged per byte. The
coordination server assigns the list of nodes in a region as a function of the
tailnet, so all nodes within a tailnet should generally be on the same node and
not require forwarding. Only after a failure do clients of a particular tailnet
get split between nodes in a region and require inter-node forwarding. But over
time it balances back out. There's also an admin-only DERP frame type to force
close the TCP connection of a particular client to force them to reconnect to
their primary if the operator wants to force things to balance out sooner.
(Using the (*derphttp.Client).ClosePeer method, as used by Tailscale's
internal rarely-used cmd/derpprune maintenance tool)
We generally run a minimum of three nodes in a region not for quorum reasons (there's no voting) but just because two is too uncomfortably few for cascading failure reasons: if you're running two nodes at 51% load (CPU, memory, etc) and then one fails, that makes the second one fail. With three or more nodes, you can run each node a bit hotter.