SBM: Storage Benchmark Monitor

SBM is SBK’s distributed measurement aggregator. SBK clients using GrpcLogger send SBP/gRPC latency records to SBM, which merges them into cluster-wide periodic and total statistics.

SBP 4.0 uses one ordered client-streaming RPC per SBK process and packed primitive latency/count fields. GrpcLogger accumulates exact frequencies in a primitive map, submits immutable batches through a bounded sender queue, obeys gRPC flow control, and waits for the final stream acknowledgment during shutdown. SBP 4.0 intentionally removes the earlier unary latency RPC and protobuf map field, so SBK and SBM must use the same SBP major version. SBM explicitly configures and advertises its inbound record limit (maxRecordSizeMB, 16 MiB by default); SBK sends no more than the smaller client/server limit.

SBM does not execute storage operations and does not launch remote processes. Use SBK for load generation and SBK-GEM for SSH orchestration.

Data flow

flowchart LR
    A[SBK client A / GrpcLogger] --> G[SBM gRPC service]
    B[SBK client B / GrpcLogger] --> G
    G --> Q[Concurrent ingestion queues]
    Q --> R[Aggregate recorder]
    R --> O[Prometheus, Local Web Console, or result output]

The default gRPC port is 9717. Container configuration also exposes the configured metrics port. Keep the gRPC service on a trusted benchmark network unless an external security layer is provided.

Build

./gradlew :sbm:check
./gradlew :sbm:installDist

The launcher is generated under sbm/build/install/sbm/bin/sbm.

Run

Display authoritative options:

./sbm/build/install/sbm/bin/sbm -help

Start with defaults:

./sbm/build/install/sbm/bin/sbm

When a positive -records N is supplied to identify a fixed-record run, SBM exits with a failed benchmark if no SBK performance batch arrives for -idletimeoutseconds N; the default is 600 seconds. Without -records, the deadline is disabled. The check runs only while the ingestion queues are empty. Every SBP batch containing completed records renews the complete deadline; an empty periodic batch is still printable but does not represent progress. The idle timeout must be strictly greater than the selected logger’s reporting interval.

SBM’s final lifecycle log distinguishes an orchestrated successful -seconds or -records completion from an explicit stop, idle timeout, or local/remote internal failure. Failures reported by remote SBK clients and failures raised while draining the final aggregation window remain visible in the returned completion and terminal log.

Start the aggregate Prometheus exporter explicitly:

./sbm/build/install/sbm/bin/sbm \
  -out SbmPrometheusLogger -class file -action r

SBM prints ready-to-copy scrape URLs and exposes aggregate metrics at http://<sbm-host>:9719/metrics by default. Register that endpoint—not each gRPC client—in the separately deployed SBK Dashboard. Use -context PORT/PATH to change the endpoint. The PrometheusLogger and SBK Dashboard guide covers dashboard deployment, endpoint registration, metrics tags, retention, networking, and troubleshooting.

Start the dependency-free SBK Local Web Console instead of Prometheus:

./sbm/build/install/sbm/bin/sbm -out SbmWebLogger -class file -action r

The Local Web Console uses plain HTTP and listens on all IPv4 interfaces at port 9720 by default. At benchmark start and completion, SBM prints run URLs for localhost, loopback, hostname, and every usable host IPv4 address. It displays aggregate SBM connection, workload, throughput, request-pressure, timeout, and latency-percentile data. Multiple SBK, SBM, and SBK-GEM WebLogger benchmarks can share the same port, with independent run URLs and browser selections. The console keeps completed graphs while a browser remains connected and exits after the configured idle timeout with neither an active publisher nor browser activity. The default idle timeout is one minute. Remote access requires suitable routing and firewall rules; because the service has no authentication or TLS, use a trusted network or SSH tunnel.

See the WebLogger guide for web console options, lifecycle, browser leases, security, and a complete distributed example.

Then point one or more installed SBK clients at it:

./build/install/sbk/bin/sbk \
  -class file -file /tmp/sbk.bin \
  -writers 1 -size 4096 -seconds 30 \
  -out GrpcLogger -sbm <sbm-host> -sbmport 9717

Use a hostname or address reachable from every SBK client. Firewalls, containers, and NAT must permit the callback path.

Internal ownership

Class	Responsibility
`io.sbm.main.SbmMain`	Executable entry point
`io.sbm.api.impl.Sbm`	Logger discovery, arguments, and benchmark construction
`SbmBenchmark`	gRPC server and aggregate-recorder lifecycle
`SbmGrpcService`	SBP registration, ordered stream validation, and latency-record endpoints
`SbmLatencyBenchmark`	Concurrent queue ingestion and window dispatch
`SbmTotalWindowLatencyPeriodicRecorder`	Periodic and total aggregate windows
`SbmPrometheusLogger`	Aggregated output and metrics
`SbmWebLogger`	Aggregated output and local live web console publication

Protocol sources are under sbk-api/src/main/proto; generated Java/gRPC sources are build products.

Operational checks

Confirm every client uses compatible SBP/protobuf definitions.
Synchronize clocks when interpreting cross-host timestamps or correlated events.
Record client count, worker count, network topology, interval, and latency bounds.
Watch rejected registrations, disconnected clients, invalid latencies, and discarded lower/upper values.
Treat sequence-gap, stream-overload, and final-acknowledgment failures as invalid benchmark runs.
Do not combine already calculated client percentiles; SBM aggregates latency records/windows before reporting percentiles.

Containers and dashboards

The module’s Jib configuration builds an sbm image and declares its service/metrics ports. Repository monitoring assets are under grafana/ and container guidance under dockers/. For persistent multi-endpoint visualization, deploy SBK Dashboard and register the SBM metrics endpoint on port 9719. For a lightweight in-memory view, use SbmWebLogger instead.