neper

Author: Ludovic HENRY ludovic.henry@qti.qualcomm.com
Date: 2026-06-17
Scope: RISC-V (riscv64/linux) support status for neper
Audience: Technical leadership, resource allocation strategy
Verification policy: Every claim is cross-referenced to a primary upstream source. Items that could not be verified against a second source are marked [NEEDS VERIFICATION].

1. Project Overview

neper is a Linux network performance benchmarking tool written in pure C99. It measures TCP and UDP throughput, request/response latency, and connection rates. A separate set of binaries (psp_stream, psp_rr, psp_crr) targets Google’s PSP transport-layer encryption protocol. The tool was developed inside Google and open-sourced under Apache-2.0. It has 326 GitHub stars and no formal release tags.

Governance: No foundation, no steering committee, no MAINTAINERS or CODEOWNERS file. All active maintainers are Google employees. The most active contributors by commit count: hi-rajat-kumar (Rajat Kumar, 27 commits), kevinGC (Kevin Krakauer, 18 commits), wdebruij (Willem de Bruijn, 14 commits). External contributors exist (George Prekas/Amazon, David Gibson/Red Hat) but hold single commits and are not maintainers.

License: Apache-2.0.

Corporate sponsors: Google exclusively. Google is a RISE Premier Member, but neper is not listed as a RISE-funded project and no RISE blog posts mention it.

Culture on new ports: The RISC-V port (PR #35, 2022) was accepted with no recorded friction. The author was a Google employee. The project’s design – pure C, no external dependencies, POSIX networking only – makes porting inherently low-effort.

2. Port History and Upstreaming Timeline

Date Event Source
2022-07-28 PR #35 submitted: “neper: add support for Linux on RISC-V with musl”, author Michael Hope (mlhx@google.com, Google) PR #35
2022-07-31 PR #35 merged by Willem de Bruijn (wdebruij, Google) PR #35

PR #35 changes (19 additions, 19 deletions across 4 files):

  • cpuinfo.c: added || defined(__riscv) to an existing three-way #ifdef that selects the flat processor: N /proc/cpuinfo parsing path (shared with aarch64 and powerpc).
  • rusage.c: replaced the HZ macro with sysconf(_SC_CLK_TCK) – a musl portability fix, not RISC-V-conditioned.
  • logging.c: removed fcloseall(), a GNU extension absent from musl – not RISC-V-conditioned.
  • thread.c: reordered pthread_create() before pthread_setaffinity_np() and replaced pthread_attr_setaffinity_np() with the post-creation variant – required because musl stubs the attr variant.

The RISC-V-specific code change is a single token: || defined(__riscv) in one #ifdef. The remaining changes are musl portability fixes that benefit all non-glibc builds but happen to be required on RISC-V musl targets.

The port is fully upstream. There is no downstream fork, no patch queue, no out-of-tree maintenance.

3. Upstream Support Tier

No formal tier or platform policy document exists. No PLATFORMS.md, no MAINTAINERS file, no stated architecture support matrix.

Evidence from CI and distribution:

Dimension amd64 arm64 riscv64
CI build Yes (Docker image published) Yes (Docker image via QEMU) No
CI test No dedicated test job for any arch No dedicated test job for any arch No
Release binary No formal releases for any arch No formal releases for any arch No
Docker image at ghcr.io Yes Yes No
Distro package No No No

The only CI workflow is .github/workflows/publish-image.yaml. Its platforms: line reads linux/amd64,linux/arm64. riscv64 is absent. There is no separate build-and-test workflow for any architecture; the CI exclusively publishes Docker images.

riscv64 is at a lower support tier than arm64 by the only measurable criterion (CI coverage), but since neither arm64 nor riscv64 receives functional test coverage, the practical gap is narrow.

4. Technical Architecture and RISC-V-Specific Subsystems

neper has no architecture-specialized compute. It is a network I/O benchmark: it opens sockets, spawns threads, measures throughput and latency. There are no JIT compilers, no cryptographic primitives, no SIMD kernels, no GC barriers.

Architecture-specific file inventory:

  • No .S assembly files for any architecture.
  • No arch/ subdirectory.
  • No SIMD intrinsics (no AVX, no NEON, no RVV).
  • No vfloat32m1_t, rvv, or riscv64 symbols anywhere in the tree (searches returned zero results).

Per-component matrix:

Component amd64 arm64 riscv64 Notes
/proc/cpuinfo parsing Full topology (physical_id, core_id, siblings, cpu_cores) Flat (processor only) Flat (processor only) – identical path to arm64 One #ifdef __riscv guard routes to the correct parser
Compute kernels None None None No compute kernels exist in neper
SIMD / ISA extensions None None None Not applicable
Assembly None None None Not applicable
NUMA topology (optional) Supported Supported Supported (Debian sid) Disabled by default (-DNO_LIBNUMA)
PSP binaries Builds; requires Google-internal kernel patch Builds; not tested Builds; not tested PSP not upstream for any arch

The riscv64 implementation is complete for what neper requires. No stubs, no TODO markers, no missing paths were found.

5. Build System, Cross-Compilation, and Toolchain

Build system: GNU Make only. No CMake, meson, or autoconf. One Makefile at the repo root.

CFLAGS (hardcoded): -std=c99 -Wall -O3 -g -D_GNU_SOURCE -DNO_LIBNUMA

Link flags: -lm -lrt -lpthread

Native build on riscv64 host:

make all

Produces 8 binaries: tcp_rr, tcp_stream, tcp_crr, udp_rr, udp_stream, psp_stream, psp_crr, psp_rr.

Cross-compilation from x86_64 (no official support documented, but functional):

CC=riscv64-linux-gnu-gcc make all

or with Clang:

CC="clang --target=riscv64-linux-gnu" make all

Toolchain requirements:

  • C99 required (-std=c99 is hardcoded). Any GCC >= 4.x or Clang >= 3.x suffices.
  • -D_GNU_SOURCE is set. glibc is the natural target. musl is supported (PR #35 resolved the blockers).
  • -DNO_LIBNUMA is hardcoded in the upstream Makefile. No NUMA library is needed.
  • epoll_pwait2() is optional: pass -DHAVE_EPOLL_PWAIT2 only if your libc provides it. glibc >= 2.35 (Debian 12 / bookworm) includes this on riscv64.
  • No minimum GCC version is documented. The fcloseall() removal in PR #35 was the only musl-blocking issue and is already resolved.

Known build issue (Issue #67, open): A floating-point static initializer in histo.c fails with gcc 7.5.0 on SLES 15 SP5 (error: initializer element is not constant). Workaround: make CC=clang. GCC toolchains for riscv64 in Debian 12 and later are not affected, but embedded or vendor RISC-V toolchains on older GCC could trigger this. [NEEDS VERIFICATION for which riscv64 GCC versions reproduce this.]

QEMU: The CI pipeline installs docker/setup-qemu-action@v3 to build the arm64 Docker image. Adding linux/riscv64 to the platforms: field in .github/workflows/publish-image.yaml is all that is technically required, subject to the distroless image gap described in Section 8.

6. Feature Coverage and Gap Analysis vs arm64 and amd64

Feature matrix:

Feature amd64 arm64 riscv64
TCP throughput (tcp_stream) Yes Yes Yes (source-level)
TCP request/response (tcp_rr, tcp_crr) Yes Yes Yes (source-level)
UDP throughput/RR (udp_stream, udp_rr) Yes Yes Yes (source-level)
PSP encrypted transport Requires Google kernel patch Requires Google kernel patch Requires Google kernel patch
NUMA-aware socket distribution Yes (optional; disabled by default) Yes (optional; disabled by default) Yes (optional; Debian sid libnuma available)
/proc/cpuinfo topology Full (threads, cores, sockets) Core count only Core count only
eBPF NUMA pinning Yes (kernel BPF JIT) Yes Yes (riscv64 BPF JIT upstream since Linux 5.1)

Functional gaps: None for the core benchmark workloads. PSP is equally non-functional on all architectures outside Google’s internal kernel.

Performance gaps: No riscv64 benchmark data exists to quantify any gap (see Section 11). neper measures network I/O, not compute throughput, so the absence of SIMD is not architecturally significant. CPU overhead for packet processing at line rate could differ, but no data is available.

Known portability issue (Issue #49, open): The %m scanf specifier in cpuinfo.c is a glibc extension not present in musl. On musl-based riscv64 targets (e.g., Alpine Linux riscv64), this may fail at runtime when parsing /proc/cpuinfo. [NEEDS VERIFICATION: whether Alpine Linux riscv64 ships a glibc-compatible musl or a strict musl that omits %m.]

Floating-point semantics: No floating-point computation in the network measurement path. histo.c uses floating-point for histogram output only. No NaN or FP semantics issues expected on riscv64 beyond Issue #67 noted above.

7. CI/CD Infrastructure

The only CI file: .github/workflows/publish-image.yaml

Trigger: push to master and workflow_dispatch.

Runner: ubuntu-latest (x86_64).

What it does: Builds a Docker image via docker buildx with QEMU for arm64 emulation, then pushes to ghcr.io/google/neper. This is a packaging step, not a build-verification or test step.

Platforms line (verbatim): platforms: linux/amd64,linux/arm64

riscv64 is absent.

CI dimension amd64 arm64 riscv64
Docker image built Yes Yes (QEMU) No
Build verification (compile-only) Implicit via Docker build Implicit via Docker build Not done
Functional tests No No No
Dedicated test runner No No No
RISE hardware runner No No No

No RISE-provided riscv64 runners are used. No external CI (Cirrus CI, GitLab CI, Jenkins) exists.

8. Distribution and Release Status

GitHub Releases: The repository has zero release tags. The GitHub releases page is empty. No binary assets have ever been published.

Docker image: ghcr.io/google/neper is published for linux/amd64 and linux/arm64. No linux/riscv64 image is published. The Docker image uses a multi-stage build with gcr.io/distroless/cc-debian12:debug as the runtime base, which does not publish a riscv64 variant. This blocks riscv64 Docker image publication regardless of build success.

Distro packages:

Distribution amd64 arm64 riscv64
Debian sid No package No package No package
Ubuntu 24.04 (noble) No package No package No package
Arch Linux RISC-V No package No package No package
PyPI Not applicable (C tool) Not applicable Not applicable

neper is not packaged in any major Linux distribution on any architecture.

To obtain a working riscv64 binary today: Clone the repository and run make all on a riscv64 host or cross-compile with CC=riscv64-linux-gnu-gcc make all. No other path exists.

9. Dependencies

Summary table:

Dependency Role riscv64 Build riscv64 Test riscv64 Release Blocking Issues
glibc (system) libc, epoll, pthread, sched Yes Yes (Debian/Fedora CI) All major distros None
libm (-lm) Math (histogram output only) Yes Yes All major distros None
librt (-lrt) clock_gettime, shared memory Yes Yes All major distros None
libpthread (-lpthread) POSIX threads Yes Yes All major distros None
libnuma (optional) NUMA socket distribution, eBPF NUMA pinning Yes (Debian sid numactl 2.0.19-1+b2) No dedicated riscv64 CI in numactl Debian sid riscv64 binary available None open. Historical: numactl PR #131 (libatomic link error, closed 2022), numactl #197 (cleanup, closed 2024). Both resolved.
Linux kernel eBPF NUMA-aware socket load balancing (optional) riscv64 BPF JIT upstream since Linux 5.1 Tested in mainline kernel CI Mainline Linux None
PSP kernel interface (psp_kernel.h) Google PSP transport encryption Requires Google-internal kernel patch Not tested on any arch Not upstream on any arch PSP setsockopt constants (TCP_PSP_TX_SPI_KEY = 1730 etc.) not in mainline Linux for any architecture. Equally blocked on x86_64, arm64, and riscv64.
GCC or Clang toolchain C compiler, linker riscv64 gcc/binutils fully supported in Debian Debian builds natively on riscv64 Available in all major distros None
distroless/cc-debian12 (container base) Runtime container image No riscv64 tag published Not tested Not published for riscv64 Blocks Docker image workflow for riscv64. Does not block native builds.

Critical dependency detail – distroless/cc-debian12: Google’s distroless base image at gcr.io/distroless/cc-debian12 publishes amd64 and arm64 only. Adding riscv64 to the Docker CI workflow requires either Google publishing a riscv64 distroless image or switching the base image to one that supports riscv64 (e.g., debian:12-slim).

No JIT, crypto, or numerics dependencies exist. The only compute in neper is socket I/O and histogram output. No dependency analysis beyond the single-level above is warranted.

11. Known Bugs and Active Issues

No riscv64-specific issues or pull requests exist in the google/neper repository. Searches across issues, PRs, and commits for “riscv” and “riscv64” returned zero results.

Portability-relevant open issues:

ID Title Status Severity Notes
#67 error: initializer element is not constant (gcc 7.5.0 on SLES 15 SP5) Open (Jul 2024) Low Floating-point static initializer in histo.c. Affects older or stricter GCC. Workaround: make CC=clang. May affect vendor riscv64 toolchains on older GCC. [NEEDS VERIFICATION for riscv64 GCC versions affected.]
#49 Non-C99 %m scanf specifier in cpuinfo.c Open (Jan 2024) Low GNU extension; fails on strict musl. Affects musl-based riscv64 builds (e.g., Alpine Linux riscv64). Not a correctness bug on glibc targets.

No correctness bugs specific to riscv64 were found.

Benchmark data: No riscv64 vs arm64 or riscv64 vs x86_64 performance measurements exist in any public source (GitHub, RISE blog, web search).

12. Objections and Upstream Blockers

Stated objections: None. The upstream repository contains no discussion of riscv64 beyond PR #35 (already merged).

Technical blockers:

  1. distroless/cc-debian12 has no riscv64 image. This blocks adding linux/riscv64 to the Docker CI workflow. Mitigation: switch the runtime base to debian:12-slim or wait for Google to publish a riscv64 distroless image. This is a blocker for Docker distribution only, not for native builds.

  2. PSP binaries require a Google-internal kernel patch. This is not a riscv64-specific blocker; PSP is equally non-functional on all architectures outside Google’s fleet. No action is needed.

  3. Issue #49 (%m specifier): A one-line fix replacing %m with a portable equivalent. Trivial. No upstream maintainer has objected to fixing it.

Organizational blockers: None. The project accepted the RISC-V port in 2022 with no objections. The project has no tier policy, no NEW_PORTS review process, and no stated architecture exclusions. Google-internal review of a CI addition PR is the only gate.

Acceptance probability for a CI addition PR: High. The technical change is minimal (add linux/riscv64 to the platforms: field, conditional on resolving the distroless base image issue). Google is a RISE Premier Member, which creates organizational alignment.

13. Investment Analysis

RISE has not funded or contributed to neper on riscv64. The RISE wheel builder does not track neper (it is not a Python package). No work is already covered.

13.1 Functional Enablement

The core tool builds and runs on riscv64 today without modification. The only functional gap for musl targets is Issue #49 (%m specifier in cpuinfo.c). On glibc-based riscv64 (Debian, Fedora, Ubuntu), there is no functional gap.

Fix for Issue #49: replace %m with strerror(errno) or %s, strerror(errno) in the format string. Estimated effort: 1 hour.

13.2 Performance Optimization

neper is a network I/O benchmark. CPU overhead is dominated by socket system calls, not compute. There are no SIMD kernels to optimize. No performance gap attributable to missing ISA extensions has been identified.

Data not available: riscv64 vs arm64 throughput/latency measurements for neper on any hardware.

13.3 CI/CD Infrastructure

The only actionable CI gap is the absence of riscv64 from the Docker image build matrix. The change is one line in .github/workflows/publish-image.yaml:

platforms: linux/amd64,linux/arm64,linux/riscv64

This requires resolving the distroless base image blocker. Options:

  • Switch runtime base from gcr.io/distroless/cc-debian12:debug to debian:12-slim for the riscv64 layer. This is a one-line Dockerfile change that may face upstream resistance on security grounds (distroless is preferred for minimal attack surface).
  • Use a multi-platform Dockerfile with a platform-conditional base: distroless for amd64/arm64, debian:12-slim for riscv64 temporarily.
  • Wait for Google to publish a riscv64 distroless image (no timeline available).

Estimated effort for the CI PR (excluding distroless negotiation): 2 person-days. Negotiating the distroless base image with maintainers is the uncertainty.

13.4 Ecosystem Enablement

Not applicable. neper has no dependent package ecosystem. It is a standalone network benchmarking binary.

13.5 Summary Table

Area Work Item Effort (person-weeks) Owner Priority
Functional Fix Issue #49: replace %m with portable equivalent in cpuinfo.c 0.1 Any contributor Low
CI/CD Add linux/riscv64 to Docker image platforms: field 0.5 RISE or Qualcomm contributor Medium
CI/CD Resolve distroless/cc-debian12 riscv64 base image (negotiate upstream or switch base) 1-2 Google distroless team (external) Medium
Functional Publish riscv64 vs arm64 vs x86_64 benchmark results on reference riscv64 hardware 1 RISE / Qualcomm Low

Total investment to achieve full riscv64 parity with arm64: approximately 2-3 person-weeks, with the primary uncertainty being Google’s timeline for distroless riscv64 support.

14. Updates

No updates yet – initial report dated 2026-06-17.

15. References