Service

Get server CPU back for applications, on hardware that performs on your traffic

We move the costliest packet processing, virtual switching, overlays, firewalling and encryption, from server CPU to SmartNICs, DPUs or FPGAs. The work starts by measuring candidate hardware on your own traffic profile with RFC 2544 methodology and tuning the software data path, and the offload runs as a pilot with acceptance criteria agreed up front.

Measured impact

What changes for your organization

  • Today

    Server CPU spent on the network

    Virtual switching, overlays and encryption run in software on the host and take cores the applications were bought for.

    With the offload

    CPU back for applications

    Packet processing moves to the NIC, SmartNIC or DPU that handles it best, and the freed cores go back to workloads.

  • Today

    Hardware chosen from datasheets

    Vendor figures come from vendor test setups, and the card or switch behaves differently once your packet sizes, flows and features run on it.

    With the offload

    A purchase backed by results

    Every candidate is measured on your traffic profile with the same method, so the money goes to the hardware that performs where you need it.

  • Today

    Results nobody can compare

    Each test run uses a different setup, and the test environment itself distorts latency and loss, so the numbers start arguments.

    With the offload

    Repeatable, comparable numbers

    A verified testbed, RFC 2544 methodology and automated reports give the same answer on every run, for engineers and for the business.

  • Today

    Acceleration as a leap of faith

    An offload project starts with a vendor promise, and its real gain is known only after the rollout.

    With the offload

    A pilot with a pass mark

    The acceptance criteria are agreed before the pilot starts, so the decision to roll out rests on measured gain.

  • Today

    Slowdowns found late

    Full performance runs are slow and run rarely, so a regression in firmware or drivers surfaces close to production.

    With the offload

    Regressions caught early

    Benchmarks run in CI/CD on every firmware and driver release, so a slowdown is caught where a fix costs least.

What we deliver

  • Your traffic profile

    Packet sizes, flow counts, protocols, overlays and encryption from your own environment, written down as the profile every candidate is measured against.

  • Benchmark on your traffic

    Throughput, latency and frame loss to RFC 2544, with TRex on a verified, NUMA-aware testbed, and automated reports for engineers and for the business.

    Read the case study →
  • Software data path tuning

    DPDK, queue sizing, CPU pinning, huge pages and thread affinity tuned first, so the offload decision starts from the best the software can do.

    Read the article →
  • Offload pilot

    The function that costs most, virtual switching, overlays, firewalling or encryption, moved to a SmartNIC or DPU and measured against the criteria agreed up front.

    Read the case study →
  • Integration with your orchestration

    The offloaded data path wired into OpenStack, Kubernetes or your SDN controller, so the cards are provisioned and managed with the rest of the estate.

    Read the case study →
  • Report and recommendation

    The measured results set against your objectives, the hardware or offload we recommend, and a benchmark suite your team keeps running in CI/CD.

    Read the article →

How it runs

A hardware choice and an offload pilot follow the same order; the pilot adds the build and the integration.

  1. Discovery

    Your workloads, traffic and hardware plans, the question the evaluation answers, and the acceptance criteria and plan.

  2. Traffic profile and testbed

    Your traffic profile written down, and the testbed built as code and verified before it measures anything.

  3. Baseline

    Today’s hardware and software data path measured, then tuned, so every gain is set against a fair start.

  4. Benchmark or pilot

    Candidate hardware benchmarked on your profile, or the chosen function offloaded to a SmartNIC or DPU and integrated.

  5. Report

    Results against the acceptance criteria, in a technical and a business version, with the recommendation.

  6. Handover

    The benchmark suite in your CI/CD, documented, and our support for the rollout for as long as you want it.

Two ways in

Most clients come with one of two questions. Both start with the same measurement; the offload adds the build.

  • Offload pilot

    Your servers spend CPU on networking or encryption. We tune the software path, move the costliest function to a SmartNIC or DPU, and measure the gain against agreed criteria.

  • Hardware benchmark

    You are choosing new NICs, SmartNICs, DPUs or switches. We measure the candidates on your traffic profile and recommend the one that performs, ready for the offload.

Proof

Why CodiLime

  • Measurement we built for vendors

    Our RFC 2544 framework takes a SmartNIC from testbed verification to automated reports in CI/CD, built for hardware vendors to prove their products.

  • Offload down to the silicon

    DPDK drivers, P4 pipelines on FPGA SmartNICs and DPU data planes on NVIDIA BlueField: we wrote the code that cut a vRouter from 18 host cores to 2.

  • Network and software in one team

    The engineers who build the offload also integrate it with OpenStack, Kubernetes and SDN controllers, as on the Tungsten Fabric and ONOS projects.

FAQ

Good to know

  • Which hardware do you evaluate?

    NICs, SmartNICs, FPGA cards, DPUs such as NVIDIA BlueField, and switches. We stay vendor-neutral: every candidate is measured with the same method on your traffic profile.

  • Do you test on our traffic or a synthetic one?

    On a profile built from yours: your packet sizes, flow counts, protocols, overlays and encryption, generated with TRex in a lab. RFC 2544 tests belong in an isolated testbed, so production stays untouched.

  • How do we know an offload is worth it?

    The pilot has acceptance criteria agreed before it starts, such as host cores freed, throughput or latency on your profile. The software data path is tuned first, so the gain is measured against the best software result.

  • Will the offloaded cards fit our orchestration?

    The pilot includes integration with what you run: OpenStack, Kubernetes or your SDN controller. On our projects the cards were provisioned and managed with the rest of the estate.

  • What do we keep after the project?

    The results and the recommendation, the testbed defined as code, and a benchmark suite your team can run in CI/CD on every new firmware, driver or hardware candidate.

Start with a data-plane assessment

Tell us what you are buying or where your server CPU goes. We write down your traffic profile, agree the acceptance criteria, and hand you a plan for a benchmark or an offload pilot.