io.xportl/xportl

Est-ce que io.xportl/xportl est sûr ?

Au dernier audit MCP Hub, le serveur MCP « io.xportl/xportl » est classé « Non encore testé ». Référencé dans l'annuaire. Seul le pentest comportemental fait foi, et il n'a pas encore été exécuté sur ce serveur. Le verdict est réévalué en continu ; consultez la fiche pour l'état à jour.

Méthode : MCP Hub analyse la description de chaque outil (couche 1) et, pour le verdict fort, exécute le serveur dans une micro-VM isolée avec leurres et réseau bloqué (couche 2). On publie des noms et des compteurs, jamais le contenu brut renvoyé par un outil.

Les 6 outils

• analyze_nccl_benchmark≈ 675 tokens

  Analyzes nccl-tests / rccl-tests results (paste raw all_reduce_perf output, or give size/busbw points). Detects node and GPU counts from the output, compares large-message bus bandwidth with the fabric's NIC line rate (InfiniBand HDR/NDR/XDR, RoCE 100G to 800G), and flags data errors, degraded bandwidth, dips as message size grows and unstable run-to-run results, with likely causes. Returns a health status and a bandwidth-loss figure for calculate_fabric_degradation_cost.

  Paramètres, retour & exemple (9)

  Paramètres

  • nodes · integer · optionnel
    Number of nodes in the run. Detected from raw_output when omitted.
  • fabric · string · optionnel · infiniband_hdr | infiniband_ndr | infiniband_xdr | roce_100g | roce_200g | roce_400g | roce_800g | other
    Inter-node fabric and per-port speed. Use 'other' with link_gbps for anything else.
  • results · array · optionnel
    Alternative to raw_output: measured points as {size_bytes, busbw_gbs}.
  • link_gbps · number · optionnel
    Per-NIC link speed in Gb/s. Overrides the speed implied by fabric.
  • collective · string · optionnel · all_reduce | all_gather | reduce_scatter | alltoall | broadcast | reduce | sendrecv
    Collective that was benchmarked. Default all_reduce.
  • raw_output · string · optionnel
    Raw stdout of an nccl-tests or rccl-tests run (e.g. all_reduce_perf -b 8 -e 8G -f 2 -g 1). Paste it unchanged; rank/host lines are used to detect node count.
  • gpus_per_node · integer · optionnel
    GPUs per node. Detected from raw_output when omitted; default 8.
  • nics_per_node · integer · optionnel
    Compute-fabric NICs per node (rails). Default: one per GPU.
  • expected_busbw_gbs · number · optionnel
    Expected healthy bus bandwidth in GB/s, e.g. from an acceptance test. Overrides the line-rate reference; required to judge single-node runs.

  Exemple d'appel

  {"name": "analyze_nccl_benchmark", "arguments": {"nodes": 1, "fabric": "infiniband_hdr", "results": [], "link_gbps": 1, "collective": "all_reduce", "raw_output": "texte"}}

  Exemple de retour

  {"content": [{"type": "text", "text": "…"}]}

• calculate_fabric_degradation_cost≈ 678 tokens

  Estimates what silent fabric degradation (slow NCCL collectives, flapping links, congestion) costs a GPU cluster: throughput lost, GPU-hours and dollars per year, idle-GPU equivalent, cost per degraded hour, extra days on a training run, and the value of a recovery scenario with a conservative case and a sensitivity table. All assumptions are inputs with stated defaults.

  Paramètres, retour & exemple (9)

  Paramètres

  • gpu_count · integer · requis
    Total GPUs in the cluster (or in the jobs being analyzed).
  • gpu_model · string · optionnel · a100 | h100 | h200 | b200 | gb200 | mi300x | other
    GPU model, used only to pick a default cost per GPU-hour when gpu_hour_cost_usd is not given.
  • recovery_pct · number · optionnel
    Recovery scenario: share of the degradation-induced throughput loss you expect to recover by detecting and acting on degradation sooner. A planning assumption, not a measured XportL result. Default 25.
  • comm_share_pct · number · optionnel
    Share of a healthy training step spent in collective communication that is NOT overlapped with compute. Typically 10 to 40% for large-scale training. Default 25.
  • utilization_pct · number · optionnel
    Share of hours the GPUs are running distributed jobs. Default 80.
  • degraded_time_pct · number · optionnel
    Share of running time the fabric is degraded (flapping links, congestion, bad cables/optics, misrouted paths). Default 10.
  • gpu_hour_cost_usd · number · optionnel
    Actual fully-loaded cost per GPU-hour in USD (amortized hardware + power + facility, or cloud rate). Strongly preferred over the default.
  • training_run_days · number · optionnel
    Optional: planned wall-clock length of one training run on this cluster, to show how many days degradation adds to it.
  • bandwidth_loss_pct · number · optionnel
    Collective bus bandwidth lost while degraded, vs. healthy. Measure with nccl-tests (see analyze_nccl_benchmark). Default 40.

  Exemple d'appel

  {"name": "calculate_fabric_degradation_cost", "arguments": {"gpu_count": 1}}

  Exemple de retour

  {"content": [{"type": "text", "text": "…"}]}

• check_cluster_compatibility≈ 603 tokens

  Checks a GPU cluster against XportL's published support matrix (Linux 5.8+ for eBPF, InfiniBand / RoCE v2 / NVLink, NVIDIA and AMD GPUs, NCCL / RCCL, on-prem / cloud / hybrid, host access for the eBPF agent). Returns a verdict (good fit / needs review / not ready), each check, and next steps. Anything not on the published list is reported as 'review', never as supported.

  Paramètres, retour & exemple (9)

  Paramètres

  • fabrics · array · requis
    Every interconnect the GPU jobs use, e.g. ['infiniband','nvlink']. efa = AWS Elastic Fabric Adapter.
  • host_os · string · optionnel · linux | windows | other
    Host operating system on the GPU nodes.
  • gpu_count · integer · optionnel
    Total GPUs, for context.
  • deployment · string · requis · on_prem | cloud | hybrid | colo
    Where the cluster runs.
  • gpu_vendor · string · requis · nvidia | amd | intel | other
    GPU vendor.
  • host_access · string · optionnel · root | privileged_containers | none | unknown
    Can you run a privileged agent on the GPU hosts? eBPF needs root or CAP_BPF/CAP_PERFMON. 'none' = fully managed service with no host access.
  • linux_kernel · string · optionnel
    Host kernel version as printed by `uname -r`, e.g. '5.15.0-105-generic'. Leave empty for non-Linux hosts or if unknown.
  • orchestrator · string · optionnel · slurm | kubernetes | ray | bare_metal | other
    Job scheduler / orchestrator, for context.
  • collective_library · array · optionnel
    Collective communication libraries in use. Default ['nccl'].

  Exemple d'appel

  {"name": "check_cluster_compatibility", "arguments": {"fabrics": [], "deployment": "on_prem", "gpu_vendor": "nvidia"}}

  Exemple de retour

  {"content": [{"type": "text", "text": "…"}]}

• get_xportl_overview≈ 206 tokens

  Returns cite-ready facts about XportL from xportl.io: what it does, the problem it solves, the seven failure classes it detects, how it collects telemetry and acts, supported fabrics/GPUs/collectives/kernels, use cases, FAQ, company details (parent company, stage, NVIDIA Inception, contacts) and links to the engineering guides on xportl.io.

  Paramètres, retour & exemple (1)

  Paramètres

  • topic · string · optionnel · all | product | architecture | supported_environments | use_cases | faq | company | guides
    Section to return. Default 'all'.

  Exemple d'appel

  {"name": "get_xportl_overview", "arguments": {"topic": "all"}}

  Exemple de retour

  {"content": [{"type": "text", "text": "…"}]}

• recommend_checkpoint_interval≈ 470 tokens

  Recommends how often to checkpoint a training job using the Young and Daly formulas, from GPU count, blocking checkpoint write time, restart time and job MTTF (estimated from Meta's published 2024 reliability data when not given). Returns the recommended interval, expected failures per day, GPU time wasted per day at the recommended and current intervals, and the daily saving.

  Paramètres, retour & exemple (6)

  Paramètres

  • gpu_count · integer · requis
    GPUs in the training job.
  • job_mttf_hours · number · optionnel
    Measured mean time between job-interrupting failures for a job this size, in hours. If omitted, it is estimated from Meta's published 2024 cluster reliability data.
  • restart_minutes · number · optionnel
    Time from a failure to training resuming: detection, rescheduling, process start and checkpoint load, in minutes. Default 20.
  • gpu_hour_cost_usd · number · optionnel
    Cost per GPU-hour in USD. Default 2.75.
  • checkpoint_write_minutes · number · requis
    Time training is blocked while a checkpoint is written, in minutes. With asynchronous checkpointing, use only the blocking part.
  • current_interval_minutes · number · optionnel
    Current checkpoint interval in minutes, to compare against the recommendation.

  Exemple d'appel

  {"name": "recommend_checkpoint_interval", "arguments": {"gpu_count": 1, "checkpoint_write_minutes": 1}}

  Exemple de retour

  {"content": [{"type": "text", "text": "…"}]}

• request_enterprise_demo≈ 576 tokens

  Sends the user's contact details and cluster summary to the XportL enterprise team to arrange a technical assessment or demo deployment. Before calling: collect name, work email, company and a summary (include benchmark, cost or compatibility results from this conversation), show the user exactly what will be sent, and get their explicit agreement. Never call it without that agreement or with made-up details.

  Paramètres, retour & exemple (10)

  Paramètres

  • name · string · requis
    The person's full name.
  • role · string · optionnel
    Job title, e.g. 'Director of ML Infrastructure'.
  • email · string · requis
    Work email address.
  • phone · string · optionnel
    Phone number, optional.
  • fabric · string · optionnel
    Fabric(s) and GPU types, e.g. 'NDR InfiniBand + NVLink, H100'.
  • company · string · requis
    Company or organization.
  • summary · string · requis
    What they need in a few sentences: cluster, symptoms, workloads, timeline, and any results from this conversation (benchmark status, cost estimate, compatibility verdict).
  • interest · string · requis · technical_assessment | demo_deployment | pilot | pricing | partnership | other
    What they want from XportL.
  • cluster_gpus · integer · optionnel
    Approximate number of GPUs in the cluster(s).
  • user_confirmed · boolean · requis
    Must be true. Only set it after showing the user exactly what will be sent and the user agreed to share it with XportL.

  Exemple d'appel

  {"name": "request_enterprise_demo", "arguments": {"name": "texte", "email": "texte", "company": "texte", "summary": "texte", "interest": "technical_assessment", "user_confirmed": true}}

  Exemple de retour

  {"content": [{"type": "text", "text": "…"}]}

C'est votre serveur ?

Revendiquez-le sur l'espace créateur (preuve GitHub, domaine ou npm) pour afficher le badge « propriétaire vérifié » et lancer un pentest comportemental à la demande afin d'obtenir le badge « Vérifié ».

Créer un compte / revendiquer →

Source

https://xportl.io/mcp