TENT High-Performance TCP

hp_tcp is a standalone TENT transport for CPU DRAM transfers over data-center TCP. Standard tcp remains the RPC-based compatibility path. The first version intentionally excludes GPU memory, TLS, WRITE striping, rail failover, transparent replay after an ambiguous WRITE and dynamic lane scheduling.

Architecture

Each worker owns one asio::io_context and one thread. Each peer has a configured number of persistent lanes. A separate sequence for each peer rotates operations across them, so interleaved traffic to other peers cannot pin a peer to one lane. Each sequence starts at that peer’s first request ID to preserve its initial lane choice. Request IDs remain globally unique. A stable hash of peer and lane selects the owner; socket state never moves between workers, and operations on a lane are FIFO. ASIO provides the event queue; process-wide task and byte admission limits bound all accepted work, including callbacks waiting in that queue.

Both ends disable Nagle’s algorithm: headers and payloads use separate writes, so a delayed ACK must not hold up a short payload on a persistent socket.

The server uses the same worker pool. Accepted sockets are assigned to workers and stored in worker-owned session sets. A global connection limit bounds live sessions; closing a session removes it immediately rather than retaining one thread per historical connection.

TENT request -> bounded admission -> owner worker -> persistent lane
             -> versioned TCP protocol -> registered remote buffer

Static multi-rail routing

rail_addresses optionally assigns persistent lanes across an ordered set of local and remote TCP addresses. Each entry must be a numeric address assigned to the local host. The current implementation accepts IPv4 addresses only. When the list is non-empty it supplies the published endpoints instead of advertise_address. Lane i uses rail i % rail_count, binds its socket to the matching local address, and keeps the existing peer-and-lane worker ownership. Both peers must configure the same non-zero rail count. A specific listener address must match the sole rail; multiple local rails require a wildcard listener.

Routing is deliberately static. Small transfers and all WRITEs stay on one persistent lane. A READ large enough to contribute at least one internal I/O step per configured rail is split into one contiguous slice per rail. The slices reuse the persistent lanes and complete as one TENT task. hp_tcp does not rebalance traffic or fail over between rails.

The current internal step is 1 MiB: with two rails, READs of at least 2 MiB are sliced; smaller READs and single-rail READs are not. Remainder bytes are assigned to the first slices, so uneven lengths still cover the request exactly. Multiple independent WRITEs can use different rails, but each WRITE stays on one lane and waits for its own remote completion ACK.

connections_per_peer is the total lane budget, not a per-rail multiplier. For example, four lanes and two rails give two lanes per rail. Worker count is independent of rail count: several lanes may share an owner worker. Increasing lanes on a single rail does not split a single READ into multiple streams.

Protocol and memory safety

Requests contain a version, opcode, request ID, registration ID, remote address and length. Responses contain the request ID, status and committed byte count. A WRITE completes only after the target has copied the full payload and returned an acknowledgement. A READ completes after the full response payload arrives.

Every registered buffer has an ID formed from a per-registry random namespace and a monotonic sequence, plus a remote permission. This prevents a stale ID from a previous server incarnation from becoming valid after restart. The target validates the ID, range and permission before access. An operation holds a lease until its final I/O callback retires; unregister hides the range from new work and waits for existing leases. Stale registration metadata causes one bounded metadata refresh and retry on the same transport. Permission and range failures are terminal.

If a WRITE request may have reached the peer but no valid acknowledgement is received, the remote outcome is unknown. That failure is terminal and is not replayed through another transport; otherwise a committed WRITE whose ACK was lost could execute twice.

Timeouts and shutdown

Resolve/connect use connect_timeout_ms. Header, payload and response progress use progress_timeout_ms on both client and server. A newly accepted connection must send its first header byte before the deadline, and every partial header or payload must continue to make progress. After a valid request completes, pure idle time on its persistent connection is not treated as stalled I/O; the deadline resumes as soon as the next header begins. A timeout cancels the resolver or socket; terminal completion is published only after the corresponding callback retires.

A failed READ slice cancels its siblings; the logical task settles only after all slice callbacks retire. A stale-registration result can trigger the existing bounded metadata refresh/retry, not migration onto another rail. Independent peers can continue while a peer is waiting for its progress timeout. FIFO sharing within one peer’s lanes can still delay small requests behind large ones; slicing is not a priority or preemption mechanism.

The client separately closes a pooled socket after idle_connection_timeout_ms without active or queued work on that lane (default 60 seconds). New work before expiry cancels this timer and reuses the socket; work after expiry reconnects. This releases receiver connection slots held by idle updated clients. The server does not evict established idle sockets: it cannot know whether a client has just started another WRITE. Older clients that keep sockets open indefinitely still require their own pool cleanup. Tasks attempted while the receiver connection limit is full can still fail; idle cleanup is not task backpressure or an automatic retry policy.

Shutdown closes admission and the listener, drains queued dispatch callbacks, cancels every client lane and server session on its owner, waits for operations and leases, then stops and joins worker threads. This makes shutdown bounded even when a peer sends only part of a request.

This ordering is a lifecycle invariant, not an incidental destructor detail: the client and server are destroyed before the worker contexts they use. In a debug build, normal teardown asserts that admission, client operations and server sessions have all drained before their owners are destroyed.

An exception escaping an ASIO handler marks the runtime failed and blocks further admission. The owner event loop continues only to retire previously committed work and process teardown cancellation with the same affinity. Once those resources drain, shutdown joins the workers and reports the failure. Likewise, admission-release underflow is fail-closed: counters are preserved, new work is rejected, and drain returns an error instead of treating live work as complete.

Configuration

The transport is configured under transports.hp_tcp:

Field	Meaning
enable	Enable hp_tcp; set transports.tcp.enable to false. The two transports cannot be enabled together because control-plane notification ownership is singular.
bind_address, advertise_address, port	Listener and published endpoint.
rail_addresses	Ordered numeric IPv4 source addresses for static lane-to-rail routing. The list must be unique, no longer than connections_per_peer, and have the same length on both peers. A non-wildcard bind_address must equal the sole rail address; multiple rails require it to be empty or 0.0.0.0.
worker_count	ASIO event-loop threads.
connections_per_peer	Persistent lanes per peer.
max_outstanding_tasks, max_outstanding_bytes	Global admission bounds.
max_transfer_bytes	Maximum request size. When HP TCP is enabled, coalescing of HP TCP/UNSPEC requests respects both local and advertised remote limits; an individually oversized request is still rejected.
connect_timeout_ms, progress_timeout_ms	Connection and I/O deadlines.
idle_connection_timeout_ms	Positive client idle-pool retention time; default 60000 ms. Active or queued requests are never expired by this timer. Shorter retention frees receiver slots sooner but requires more reconnections for intermittent traffic.

Single-rail and paired-rail examples

For a single rail, put this in the server’s MC_TENT_CONF JSON file (replace the example address with an address assigned to the host):

{
  "transports": {
    "tcp": {"enable": false},
    "rdma": {"enable": false},
    "shm": {"enable": false},
    "hp_tcp": {
      "enable": true,
      "bind_address": "",
      "rail_addresses": ["10.0.0.2"],
      "worker_count": 4,
      "connections_per_peer": 4
    }
  }
}

Use the same configuration on the client with its local address 10.0.0.1. For two rails, change only the lists:

Host	rail_addresses
Client	["10.0.0.1", "10.1.0.1"]
Server	["10.0.0.2", "10.1.0.2"]

Entries pair by index, and both hosts need working source-address routes for those pairs. Reused sockets retain that mapping; failed sockets are closed before later requests reconnect.

MC_TENT_CONF loads a complete configuration, so include the transport enable flags even when using tebench’s --xport_type=hp_tcp. To check data with the existing benchmark, start its target, then run an initiator with its advertised segment name:

MC_TENT_CONF=client.json tebench --backend=tent --xport_type=hp_tcp \
  --tent_transport_hint=hp_tcp --target_seg_name=SERVER_SEGMENT \
  --seg_type=DRAM --op_type=mix --check_consistency=true \
  --start_block_size=67108864 --max_block_size=67108864 --duration=3

Repeat with both block-size flags set to 4096 for the unsliced path. With --xport_type=hp_tcp --check_consistency=true, the CPU checker uses seed-reproducible, non-constant data and a full byte comparison to detect reordered slices. Run throughput separately without this checking overhead. Ordinary mix, other backends and write_seed/read_verify retain their existing data patterns.

Checking rail use

Test-local socket relays record the peer/local addresses and completed slice ranges. Full-engine tests check payloads, guards, the slicing threshold and connection reuse; two-process E2Es also cover unequal transfer-size limits.

On two machines, inspect connections with ss -tnp, source-address routes with ip route get REMOTE from LOCAL, and per-interface byte counters before/after a transfer. READ payload moves from server TX to client RX. Compare both rails’ deltas with successful application bytes; account for protocol overhead and unrelated traffic. Two open connections alone do not prove payload use.

Loopback proves routing, not physical NIC use. Check PCI devices and shared host/fabric limits. Compare one rail/one lane, one rail/multiple lanes, and two rails/the same total lanes: the last two differ in single-READ slicing as well as rail placement.

Tuning concurrent READs

For one peer with C outstanding large READs and R rails, at most min(connections_per_peer, C * R) payload streams can be active. This assumes the READs are large enough to slice across all rails. Four outstanding READs with four total lanes therefore have the same four-stream upper bound with one or two rails; adding a rail alone does not increase that bound.

With four rails and four total lanes, each rail has one connection. Eight lanes provide two per rail. Each lane processes one operation at a time; once those four lanes already have distinct owners, adding workers alone cannot increase the number of active lane owners. Compare lane and worker counts separately, recording the counts at both endpoints, throughput, tail latency and CPU use. Additional lanes sharing the same workers need not increase throughput.

Keep CPU and memory placement fixed during these comparisons; record the NICs’ NUMA nodes. Use per-thread CPU measurements because Store and tebench callers also consume CPU polling for completion. A stream-count upper bound does not predict throughput or establish a universal rail/worker default.

Measured scope

On two Xeon 8457C virtual machines, with four workers and four total lanes, three interleaved runs (1-second warmup, 3-second measurement) gave the following medians:

READ workload	Metric	One rail	Two rails
64 MiB, one concurrent task	Throughput (GB/s)	3.22	6.46
64 MiB, four concurrent tasks	Throughput (GB/s)	11.13	10.96
4 KiB, one concurrent task	Mean latency (microseconds)	72	78

Both interfaces carried payload-direction traffic, but their underlying resource independence is not guaranteed. In that four-lane/four-worker setup, four concurrent READs showed no additional gain. A same-pool 4 KiB/64 MiB closed-loop mix still delayed small tasks behind large ones: static slicing offers neither latency isolation nor universal bandwidth scaling.

A Store get_into comparison on two H20 hosts used the unchanged 202ad9c89 Release build, 8 MiB host-memory objects and four closed-loop callers (one outstanding READ each). Both endpoints used the worker/lane counts below, CPU 0-89 and new allocations bound to NUMA node 0. Two rails used eth1/2 on node 0; four rails also used eth3/4 on node 1. There were three 30-second runs per configuration after a 2-second warmup. Each pair of configurations was interleaved; the three pairs ran sequentially.

Rails	Workers / total lanes	Median GB/s [min, max]	Client CPU (core equivalents)
2	4 / 4	10.371 [9.664, 10.747]	6.84
2	8 / 8	8.022 [7.904, 8.159]	5.40
4	4 / 4	7.599 [7.440, 7.877]	6.86
4	4 / 8	8.537 [8.139, 8.624]	6.93
4	8 / 4	7.339 [7.260, 7.790]	6.83
4	8 / 8	11.487 [11.343, 11.551]	7.76

For four rails, increasing both counts improved throughput by 51.2% over four/four, with client CPU rising from 6.86 to 7.76 cores. Increasing workers alone did not help; increasing lanes alone helped less. The four-rail 8/8 configuration was 10.8% faster than the best measured two-rail configuration (4/4), with 13.5% more client CPU. Server-process CPU samples also rose from about 2.24 to 3.35 core equivalents. Two rails regressed with 8/8, so these results do not justify raising defaults.

Before each of the 18 timing runs, all 32 objects passed bytewise checks. Every run used the configured 4 or 8 connections, and per-rail payload byte totals matched the static split. Separate stack samples confirmed four active receive workers in the four-lane case and were excluded from timing results. The experiment changes both endpoints’ worker counts; it does not isolate client versus server costs or remove NUMA effects. It covers Store host-memory READs, not GPU transfers or model-level performance.