@mark1russell7/harness 0.2.0 → 0.4.0

package/README.md

@@ -8,33 +8,66 @@ The harness does not compute predictions. The author of a cell computes the pred
 
 A **cell** (`cell.ts`) is all that one measurement needs, as data:
 
-- **The topology.** A stock target image, pinned by digest.
+- **The topology.** A stock target image, pinned by digest, and the hops between the client and the target (`Cell.topology`). Without hops, the client and the target are a direct pair.
 - **The settings.** The sysctls of the client namespace, and the socket options of the client.
-- **The workload.** A connection that stays idle, or a connection with a request in flight.
-- **The fault.** A black hole, a reset, a frozen process, or a stall.
-- **The prediction.** An interval [T_lo, T_hi] after the event that arms the timer, or silence in a window. After a stall, the prediction also tells when the connection recovers.
+- **The workload.** A connection that stays idle, a connection with a request in flight, or a stream of requests. A stream of `GET` requests for a long value keeps many segments in flight.
+- **The fault.** A black hole, a reset, a frozen process, or a stall. In a topology, also a fault at a router or a proxy, or a router that forgets a flow. Also a proxy that closes an idle connection, or a toxic of Toxiproxy.
+- **The prediction.** An interval [T_lo, T_hi] after the event that arms the timer, or silence in a window. After a stall, the prediction also tells when the connection recovers. A silent prediction can also tell when the reply comes back (`replies`).
+- **The assumptions of the band.** A band can name the range of first round trips that it assumes (`rtt`), and the proxy times that it includes (`through`).
 
 ## The environment
 
-`openLab(target)` (`lab.ts`) builds one environment with Testcontainers:
+`openLab(target, client, topology)` (`lab.ts`) builds one environment with Testcontainers:
 
-- A new network and the target.
+- A chain of new networks, and the target on the last network. A direct pair has one network.
 - A client container that holds its network namespace.
-- A privileged sidecar in the namespace of the client, for its sysctls. Docker Desktop keeps `/proc/sys` read-only in a container that operates.
+- A privileged sidecar in the namespace of the client, for its sysctls and its routes. Docker Desktop keeps `/proc/sys` read-only in a container that operates.
+- Each hop of the topology, on the network of each of its sides.
 - An injector sidecar in the namespace of the target, with NET_ADMIN. Thus the image of the target stays stock.
-- The prober ([`@mark1russell7/prober`](https://www.npmjs.com/package/@mark1russell7/prober)): a privileged container that loads the kernel program and watches the two namespaces. Each run keeps its records from its start to its end. A run whose records the prober lost is void.
+- The prober ([`@mark1russell7/prober`](https://www.npmjs.com/package/@mark1russell7/prober)): a privileged container that loads the kernel program. It watches the namespace of the client, of the target and of each hop. Each run keeps its records from its start to its end. A run whose records the prober lost is void.
 
 Each command is an argv array.
 
+## The topologies
+
+A topology (`Topology`) is the chain of hops from the client to the target, as data. Each hop is on two networks, one for each side. Network j joins element j and element j + 1: the client is element 0, and the target is the last element.
+
+- **A router** (`RouterHop`) is the router image of the harness. It forwards, and it translates the addresses of the side of the client (NAT) with conntrack. Its idle timeout is `nf_conntrack_tcp_timeout_established`, in whole seconds. The router is privileged, thus it sets the sysctls of its own namespace, and the `sysctls` of its hop. An injector sidecar shares its namespace. Its form (`MiddleboxForm`) tells what a late packet meets after the router forgets the flow:
+
+  | Form | At the expiry | A late packet from inside | A late packet from outside |
+  | --- | --- | --- | --- |
+  | `reset-both` | a reset to the two ends | dropped | dropped |
+  | `reset-next` | nothing | a reset to its sender | dropped |
+  | `silent-drop` | nothing | dropped | dropped |
+  | `recreate` | nothing | passes, and the flow starts again | dropped |
+  | `none` | nothing | passes, and the flow starts again | a reset from the stack of the router |
+
+- **A proxy** (`ProxyHop`) is a stock image, pinned by digest, that ends each TCP connection and opens its own to the next hop. Its files go into it before it starts, and `{{next}}` in them is the address and port of the next hop. An injector sidecar shares its namespace. A privileged sidecar sets its sysctls when the hop names them.
+
+The element before a router reaches the networks after the router through it. Its privileged sidecar, or the router itself, adds each route. A router translates the address of the side of the client, thus the target and each later hop see the router as the peer.
+
+A lookup treats an expired conntrack entry as gone, thus the router forgets a flow at its timeout, to the jiffy. The DESTROY event of the entry waits for the garbage collector or a list of the table. Thus `ctwatch.py` in the router lists the table each `pump` (`PUMP`, 0.1 s by default). A DESTROY then follows an expiry within the pump and two lists of the table.
+
+`LISTING` (20 ms) is the longest list of the table that a prediction assumes. A longer list writes a `pump-slow` line, and a run with one near a detection by a timer of conntrack is void. The `reset-both` form sends its resets on the DESTROY event. Each reset has the sequence number that its end expects. A raw sniffer reads it from the last acknowledgment of that end.
+
+The lines of each router (`Lab.middlebox`) are the conntrack events and the resets that it sent, on CLOCK_MONOTONIC. Each run keeps the lines from its start to its end.
+
+## The clients and the agents
+
+- **A client image** (`ClientImage`, `Cell.client`) replaces the TCP client of the harness in each run. It operates in the namespace of the client. Its command gets the arguments of the TCP client, and it prints the same JSON lines.
+- **An agent** (`lab.agent(image, command, { pid })`) is a container in the network namespace of the target. With `pid`, it also shares the process namespace of the target, thus it can signal the server process. Its JSON lines and its `exec` stay available to the caller.
+
 ## The images
 
-The harness builds two images. The tag of each image comes from the content of all the files that its Dockerfile reads. The harness copies these files into a new build context, from its own package and from the prober package. Thus the build works in a workspace and from npm.
+The harness builds three images. The tag of each image comes from the content of all the files that its Dockerfile reads. The harness copies these files into a new build context, from its own package and from the prober package. Thus the build works in a workspace and from npm.
 
 - **The harness image** (`image/Dockerfile`) has two stages, each a digest-pinned Alpine. `Cargo.lock` pins the crates. The image holds:
   - the kernel program of the prober, compiled against the BTF of the kernel that builds it, and its Rust collector.
   - `client`, the TCP client of a cell. Rust's `std::net` sets no socket option by itself, thus the measurement gets the timers of the namespace. The client also samples `TCP_INFO` each 2 ms, a second vantage on its socket. It waits in `poll()`, which does not lock the socket. A keepalive timer that finds its socket held tries again HZ/20 later.
-  - `inject`, which applies a fault in one `iptables-restore` commit, with monotonic stamps around the commit. Thus the start of a fault is known in the VM. A `docker exec` from Windows comes 190 ms late at the median.
+  - In a stream, a reply of the `client` is whole at its CRLF, or after the length of a bulk string. The client then writes a `replied` line.
+  - `inject`, which applies a fault in one `iptables-restore` commit, with monotonic stamps around the commit. Thus the start of a fault is known in the VM. A `docker exec` from Windows comes 190 ms late at the median. With `--via DEV`, the rules go in the FORWARD chain, on the device toward one side.
 - **The Node client image** (`image/node`) is a driver client: ioredis 6.0.0 on Node 24.21.0. The digest of the image and an npm lockfile pin them. `client-ioredis.mjs` prints the lines of the TCP client on the same clock, and it adds the status changes of ioredis as `status` lines.
+- **The router image** (`image/router`) is a digest-pinned Alpine with conntrack-tools, nftables, iproute2 and Python. `mbx.sh` sets the form of the router in one nftables table, `ctwatch.py` writes the conntrack events, and `resetter.py` sends the resets of the `reset-both` form. Each run records its packages in the manifest.
 
 The build does the `cargo test` of the two crates. Thus the record decoder passes its tests before an image exists. Each amount that the native programs print is a quantity in a canonical unit. For example, an RTO is in `jiffy`, a round trip in `us`, and a segment in `B`.
 
@@ -50,10 +83,22 @@ ioredis sets the keepalive of its socket itself, after it connects. Thus a cell
 - **A simplex partition** (`simplex`, `Cell.drops`) drops the segments to the server, or the segments to the client, but not both.
 - **A FIN** (`fin`) stops the process of the peer with SIGKILL while its kernel stays up. For the client, the target starts its server as the child of a shell (`Target.command`), and `Target.kill` stops it. For the server, the harness stops the container of the client.
 - **An MTU black hole** (`mtu-blackhole`, `Cell.mtu`) drops each segment longer than the length, in the namespace that receives it.
-- **netem** (`netem`, `Cell.netem`) is tc netem on the egress of the target: loss, duplication, reordering or corruption. The outcome is a distribution.
+- **netem** (`netem`, `Cell.netem`) is tc netem on the egress of the target: loss, duplication, reordering or corruption. The outcome is a distribution. At a hop, netem is on the egress of the device of the hop toward one side.
 - **A latency** (`Cell.latency`) is tc netem delay on the egress of the target before the client connects. Thus the round trip and the RTO include it.
-
-A length match and netem act on GSO super-packets of up to 64 KB, not on segments, while TSO, GSO and GRO are on. Thus the cells of `mtu-blackhole` and `netem` turn them off on the two ends (`lab.offload`).
+- **A fault at a proxy** (`Cell.faultAt`) puts the rules of a fault in the namespace of the proxy, toward the client or toward the target. The rules act on the sockets of the proxy.
+- **A fault at a router** (`Cell.faultAt`) puts FORWARD rules in the namespace of the router, through its injector sidecar. A router forwards past INPUT and OUTPUT. Each rule faces one side by the device toward it, and by the address of the peer on that side:
+  - A black hole drops what comes in from that side (`-i DEV -s PEER`) and what goes out to it (`-o DEV -d PEER`). A simplex partition drops one of the two.
+  - A reset answers each segment from that side with a reset from the router. The reset has the address that the segment went to.
+  - An ICMP unreachable comes from the address of the router on that device, not from the peer. Its rate limit is in the `sysctls` of the router.
+  - The rules of the fault are in the filter table of iptables. The form of the router is a table of nftables at the same priority. Thus the two can both drop a packet, and an accept of one does not pass a drop of the other.
+  - Conntrack sees each packet before a FORWARD rule drops it. Thus a retransmission into a black hole at the router refreshes the flow. The resets of the `reset-both` form leave from the stack of the router, past its own FORWARD rules.
+- **No fault** (`none`) injects nothing. A calibration cell measures the path itself (`Cell.calibrates`).
+- **A forgotten flow** (`middlebox-forget`) injects nothing. The router of the topology forgets an idle flow at its own timeout. `Cell.idleBefore` holds an in-flight request until after the forget.
+- **An idle proxy** (`proxy-idle`) injects nothing. The proxy closes an idle connection at its own timeout.
+- **A toxic** (`toxic`, `Cell.toxic`) is one toxic of Toxiproxy, added through its HTTP API by `inject http`, with monotonic stamps around the request. Its name is `fault`, and the harness removes it after the run. A cell whose fault is another one adds its toxic first: a slow close, then the close of the server.
+- **A close of the server** (`server-close`) makes the server close the connection of each client, with `Target.close`: `CLIENT KILL` for Redis.
+
+A length match and netem act on GSO super-packets of up to 64 KB, not on segments, while TSO, GSO and GRO are on. Thus the cells of `mtu-blackhole` and `netem` turn them off on the two ends (`lab.offload`). They also turn them off on the two devices of a hop where the fault is.
 
 Each run starts from kernels that keep nothing from the run before it (`lab.forget()`). When a connection closes, each kernel keeps TCP metrics for the address of its peer. These metrics set the first round trip of the next connection, and thus its RTO. The harness removes the `tcp_metrics` of the two namespaces before each run.
 
@@ -61,14 +106,21 @@ Each run starts from kernels that keep nothing from the run before it (`lab.forg
 
 `measure(cell, run)` (`observe.ts`) reads a run from the records of the kernel:
 
-- **The socket.** The socket of the run is the socket that the client connected from, in its namespace.
+- **The socket.** The socket of the run is the socket of its observer. For the client, it is the socket that the client connected from, in its namespace. In a topology, the address of a hop serves each run. Thus the socket of the target or of a proxy is the socket that became ESTABLISHED in the run (`socketOf`).
 - **The time.** It times the detection from the reference event of the prediction to the error report of the kernel. eBPF stamps the two events. For keepalive, the reference is the last received segment. For the give-up, it is the first retransmission of the RTO timer, not the tail loss probe one RTO before it. For a reset, it is the send of the client, which no kernel record marks. The client stamps the send before its write.
 - **The other views.** It also gives the same interval from the `TCP_INFO` samples of the client, and the delay before the client saw the error. It also gives the steps of the kernel: each keepalive probe, or each expiry of the RTO timer.
 - **Void runs.** A run is void when the timer that drives the measurement was two jiffies or more behind its wheel slot in the timed interval. The wheel does a slot in the tick when it is due. Thus such a firing measured the host, which stopped a vCPU, not the kernel. A keepalive timer that found its socket held also makes the run void. A run whose fault came too late is void too.
-- **Findings.** A finding is a wrong errno, a detector that did not fire, or a detection where the prediction is silence.
+- **The first round trip.** The measurement gives the first round trip of the socket of the observer (`rtt`). It is the smoothed round trip in the first record of the socket: the sample of the handshake. A run whose first round trip is above the range of its prediction is void, with the reason `rtt-over-assumption` and the value. Such a run measured another path than the band assumes. For example, a round trip of 4.826 ms makes an RTO of 52 jiffies, and a ladder that assumes 1 ms has 51.
+- **Findings.** A finding is a wrong errno, a detector that did not fire, or a detection where the prediction is silence. A router that forgot the flow in the window of a silent prediction is also a finding.
+- **A reply.** A silent prediction with `replies` times the send of the request to the first segment with data that comes back. A proxy can acknowledge the request first, with no data.
+- **A toxic.** A toxic counts from the start of its commit (`fault`), as the injector stamps it.
+- **A topology.** A run in a topology also gives the bytes that the client read after its request, and when a router forgot its flow (`forgot`).
 - **Recovery.** After a stall (`prediction.recovers`), it times the send to the first segment back after the stall stopped. That segment is the answer of the server to the retransmission that got through. A run whose stall stopped outside the window of its prediction is void. A run where nothing came back is a finding.
 - **An end of file.** A FIN counts from the change of state of the socket of the peer when it closes (`peer-closed`). It ends at the change of the socket of the observer to CLOSE_WAIT. The kernel stamps the two changes.
-- **A stream.** In a stream (`probe.mode` is `stream`), the client sends `count` requests. The run keeps the time from each send to the first data after it (`samples`).
+- **A stream.** In a stream (`probe.mode` is `stream`), the client sends `count` requests. The run keeps the time from each send to the last byte of its reply (`samples`). A record without `replied` lines gives the time to the first data.
+- **Reordering.** A measurement-only run also gives what reordering, duplicates and losses did (`reordering`). At the client, it counts the arrivals beyond a hole and the arrivals of data that it had already, against `rcv_nxt`. From the sender, it gives the most SACKed and lost segments, the DSACKs, the fast retransmissions, and the reordering degree and count of the kernel.
+- **An arrival is not always one segment.** The segments that arrive while the owner holds the socket wait in its backlog, and the kernel joins the contiguous ones. Thus the kernel counts out-of-order segments (`rcv_ooopack`), and the prober counts arrivals.
+- **A proxy's own time.** A calibration cell (`Cell.calibrates`) gives the time of its proxy for each request and its reply (`crossings`). Each time goes from an arrival on one socket of the proxy to its next write on the other socket (a `tx` record).
 
 ## The verdict
 
@@ -78,18 +130,23 @@ Each run starts from kernels that keep nothing from the run before it (`lab.forg
 - **Median equivalence.** The order-statistic interval of the median must be in the band, made 5 % wider.
 - **Extension.** A median across the edge of the band extends the cell from 11 runs to 22, then 44. Then the verdict is inconclusive.
 - **Recovery.** The recovery after a stall gets the same test as a detection, by its band. A detection is then a finding.
-- **Measurements.** A cell whose outcome is a distribution (`detects: null`) has no band. Its decision is `measured`, and the verdict gives the median and the slowest reply time.
+- **Measurements.** A cell whose outcome is a distribution (`detects: null`) has no band. Its decision is `measured`, and the verdict gives the median, the 99th percentile and the slowest reply time. It adds the sum of what reordering did, when anything came out of order, twice or SACKed.
+- **A reference.** A measurement-only cell that names a reference (`from`) gives the count of each outcome. It also gives the median and the range of the time from the reference. The time ends at the end of the connection, or at the last data.
+- **A calibration.** A calibration cell gives the spread of the time of its proxy in each direction (`proxy`). The spread is the count, the median, the 99th and 99.9th percentiles by nearest rank, and the maximum.
+- **A reply.** A reply in its band (`replies`) gets the same test as a detection.
+- **The time of a proxy.** A band through a proxy includes an upper quantile of the time of the proxy for each crossing (`through`). The author of the cell adds it from a calibration cell, and the verdict tests the band as it is.
 
 The distribution test needs a simulator, which does not exist at this time.
 
 ## Runs
 
 - **`runCell(cell)`** (`run.ts`) divides the 11 runs of a cell across two new environments. It records the manifest of each environment, and it removes the fault after each run.
+  - Each environment does one warm-up before its counted runs. On a loaded machine, a new environment is sometimes slow at first. A reset cell once measured 23 ms against a median of 0.3 ms. The result keeps the warm-ups apart (`warmups`), and the verdict does not count them. A cell whose fault ends its environment, a client `fin`, does a dry warm-up without the fault.
   - The `concurrency` of a cell sets how many environments operate at the same time, each with its own prober.
   - The manifest holds the kernel release, the architecture, the boot id and the Docker version. It also holds the image ids, the packages of the image, the client and the sysctls as read back.
   - A cell whose probe names a `driver` opens its labs with the client of that driver (`openLab(target, "ioredis")`).
   - The harness sets the `targetSettings` of a cell in the target before a client connects, with the `settings` commands of the target (`redis-cli CONFIG SET`, then `CONFIG GET`).
-- **The observer.** A cell observes the client, or the server (`observer`). A cell on the server side measures the socket of the target for the connection of the client. The run waits for the error of that socket. `freeze-process` freezes the peer of the observer: the target for the client, the client for the server.
+- **The observer.** A cell observes the client, the server, or a socket of a proxy toward one side (`observer`). A cell on the server side measures the socket of the target for the connection of the client. The run waits for the error of that socket. `freeze-process` freezes the peer of the observer: the target for the client, the client for the server.
   - A cell on the server side sets the sysctls of the target through `targetSysctls`, with a privileged sidecar in the namespace of the target. That sidecar starts only when a cell asks for it.
   - Its in-flight data is the reply of the server. Thus the client sends `probe.request` first, for example a BLPOP with a timeout, and the fault occurs while the server holds it.
 - **`judge(cell, runs)`** (`run.ts`) measures and judges the raw records of a cell with the code at this time. `runCell` uses it during a cell. A holder of raw records can judge them again after a change to `observe.ts` or to the verdict.
@@ -107,6 +164,6 @@ A run needs Docker, and privileged containers for the prober and the sidecars. T
 
 `pnpm --filter @mark1russell7/harness test` starts the tests. They measure test records with the shape of real kernel records (keepalive, and the tail loss probe before the retransmissions of the RTO timer). They also test the containment, the median equivalence, the extension and the silence of the verdict.
 
-The cells themselves need Docker and privileged containers. CI does not start them.
+The cells themselves need Docker and privileged containers. CI does not start them. `HARNESS_LAB=1` starts the lab tests on Docker: a client image, an agent, and a reset and netem at a router.
 
 `pnpm --filter @mark1russell7/harness build` writes the package for npm into `build/`.