Overview
Summary
Ouroboros-Network
We released network-mux-0.12.0.0, ouroboros-network-1.3.0.0 and
cardano-diffusion-1.3.0.0, followed by the ouroboros-network
2.0 release, which dmq-node has already adopted.
cardano-diffusion 1.1.0.0 and 1.1.1.0 were released
earlier.
Tx-submission. The V2 tx-submission without a central decision
thread was merged. Since then we fixed duplicate txids in the
outbound side, fixed truncated duration counters (waits
shorter than a millisecond were counted as zero), switched outbound
calculations to Int, and made awaitSharedChange an STM
action so dmq-node can resolve races in a single transaction.
NodeToNodeV_16 with configurable tx-submission limits was
merged. Bounding the work done during a sweep (at most 1000
entries per STM transaction) is in progress.
Network-mux. Mux SDU header encoding/decoding is now about 70%
faster and allocates over 90% less. The mux now forces the first 256 KiB of a
message in the sender's thread, so an unevaluated encoding waiting in
the egress buffer no longer retains whatever it refers to. Mux
counters were added, enabling per-mini-protocol dashboards. Fixes to
mux timeouts (cancelling timeouts when actions throw and keeping
timeout monitors cancellable under masking) are in review, and an RTT-aware
mux using echo-cookie RTT sampling is being prototyped. Work has
started on a mux egress scheduler, which serves bearers by tiers
and tracks egress counters and busy periods, alongside batching changes in the
mux (withBatch, GrantAt).
Connection manager & inbound governor. Several race conditions were
fixed: connection-manager / inbound-governor races (including a
connection leak), a race on new connections in the inbound governor
(with a new IOSimPOR test), and the inbound governor no longer blocks on a
stale MuxFinished. Two small memory leaks were fixed, in the
KeepAlive server and in cardano-node's
acceptAcceptor. The diffusion layer now filters configured
or negotiated mini-protocols before passing them to the mux.
Peer selection. Peers whose peer-share request failed are now backed
off (default 1 hour), and the peer-share time is only updated for
peers we are still connected to. Topology reload of local roots
given as IP addresses is fixed (previously they waited for DNS-named
local roots, never updating when there were none). DNS TTL handling was
fixed, and upstreamyness tracing was added, also reducing
the likelihood of immediately re-promoting hot peers. We also added a
Hashable API, and work on running diffusion on multiple network
interfaces is in review.
Ping. The ping API for validating SRV records / registration
certificates was merged, to be used by cardano-cli, together with
output formatting improvements; further output
improvements are in review.
Window-stats. A new window-stats library provides efficient
rolling-window statistics backed by a finger tree — a first step towards
exposing connection metrics to applications. It already powers the Leios
percentile metrics in cardano-node. Follow-ups include
API changes, renaming and an insert performance
fix.
Smol-world. We are building smol-world, a discrete-event
simulator with a realistic transport model (loss, RFC 6298 RTO, CUBIC
congestion control per RFC 9438) on a small-world topology with per-node egress
contention, to study diffusion at scale. TCP model refinements are
in review.
Leios
The improved ΔQ analysis of Linear Leios EB diffusion was
merged. It fixes three bugs in the earlier model, sweeps over EB closure
sizes to determine practical limits, and analyses the sensitivity to mempool
fragmentation (1-hop vs. multi-hop closure fetch, TxCache miss rate). We
also published the incremental decoding benchmark results in
the Leios design document. In cardano-node we added CDFs of EB
announcement delays (with a follow-up fix) and
fixed the Praos CDF; replacing the CDFs with p90/p95/p99
percentiles is in review.
We are also experimenting with a whole-closure store-and-forward EB fetch
policy with bounded hedging in ouroboros-consensus: a peer
offers a closure only once it holds all of it, receivers request only the
missing transactions, and RTT-aware deadlines trigger hedged requests to
alternative suppliers when progress stalls. In a 64-node cardano-ignite
experiment it reduced median time for a closure to reach all nodes from
5.6 s to 4.1 s, and cut redundant closure payload from 86 MB to 1 MB
(amplification 1.53× → 1.01×).
DMQ-Node
We released dmq-node-0.7.0.0 (ouroboros-consensus-4.0,
ouroboros-network-1.2, GHC-9.14 support) and
dmq-node-0.7.2.0, and updated to ouroboros-network-2.0 and
ouroboros-consensus-5.2. Integration for Node release
11.1 was completed. Other changes include peer-sharing timeouts
and an explicit CBOR decode error for local message notification,
changes to the signature outbound side, lowering the minimum
signature body size to 90 bytes, an internal --min-sig-delay
flag, a custom validation config fix, and
running Hydra CI on aarch64-linux.
IOSim, io-classes & typed-protocols
io-classes-1.11 and typed-protocols-1.2.2.0 were released.
An io-sim performance fix (reverting eager evaluation of
say/traceM/traceSTM arguments) was merged, and ppSayTrace
was added. A fix for timeout raising a stale exception after it
returned and a TimeDuration type class are in
review. IOSimPOR gained finer control over which threads take part in
schedule exploration (explorationRacyByLabel, with labelled
fork/async variants); this is in review.
Ouroboros-Network
Leios
DMQ-Node
Cardano-Node
IOSim, io-classes & typed-protocols