{ "api_version": "1", "kind": "dse_post", "self": "https://update.dsesecurity.com/api/v1/posts/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/", "item": { "id": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/", "slug": "dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone", "url": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/", "alternate_urls": { "markdown": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone.md", "json": "https://update.dsesecurity.com/api/v1/posts/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/" }, "title": "Size an MPI transport experiment by connection count, not node count alone", "summary": "What should be calculated before applying Azure HPC guidance for smaller versus larger MPI jobs?", "format": { "slug": "guide", "name": "Guide" }, "priority": { "slug": "info", "name": "Information" }, "featured": false, "image": { "theme": "continuity-recovery", "label": "Continuity & recovery", "alt": "Paired infrastructure paths converging on a stable recovered service.", "card_url": "https://update.dsesecurity.com/assets/editorial/continuity-recovery-card.webp?v=1.8.20", "hero_url": "https://update.dsesecurity.com/assets/editorial/continuity-recovery-hero.webp?v=1.8.20", "social_url": "https://update.dsesecurity.com/assets/editorial/continuity-recovery-social-v2.jpg?v=1.8.20", "width": 2400, "height": 1350 }, "topics": [ { "slug": "business-continuity", "name": "Business Continuity", "url": "https://update.dsesecurity.com/topic/business-continuity/" }, { "slug": "it", "name": "IT", "url": "https://update.dsesecurity.com/topic/it/" } ], "author": { "name": "DSE Security Editorial Team", "url": "https://update.dsesecurity.com/#editorial-team", "type": "Organization" }, "publisher": { "name": "Detection Systems & Engineering", "url": "https://dsesecurity.com/" }, "published_at": "2026-09-10T00:31:20+00:00", "modified_at": "2026-09-10T00:32:00+00:00", "reviewed_on": "2026-09-09", "reading_minutes": 2, "word_count": 237, "potentially_affected": "MPI workloads on Azure HPC VMs evaluating the source's transport-scaling guidance.", "dse_recommendation": "Record the estimated connection count alongside the proposed MPI transport experiment.", "primary_source": { "name": "Scaling HPC applications - Azure Virtual Machines - Azure Virtual Machines | Microsoft Learn", "url": "https://learn.microsoft.com/en-us/azure/virtual-machines/compiling-scaling-applications", "published_on": null, "authority": "Microsoft Learn" }, "publishing_principles": "https://update.dsesecurity.com/updates/dse-updates-editorial-methodology/", "usage_info": "https://update.dsesecurity.com/usage/", "copyright_notice": "Copyright © 2026 Detection Systems & Engineering. All rights reserved.", "content_html": "

Source facts

\n

Microsoft’s Azure HPC guidance estimates an MPI job’s maximum connections by multiplying processes per node by the square of the job’s node count. It suggests UCX_TLS=rc,sm for jobs below 256K connections and UCX_TLS=dc,sm above 256K connections. The source also emphasizes workload-specific tuning experiments rather than assuming an optimal configuration from VM selection alone. Microsoft Learn.

\n

Applicability

\n

Use this planning check when scaling an MPI job beyond an earlier test. Record both the process layout and node count; a node-count-only label does not express the source’s connection estimate. The documented smaller/larger examples do not specify the exact boundary case.

\n

DSE recommendation

\n

Record the estimated connection count alongside the proposed MPI transport experiment. Have the application owner compare the intended job layout with the transport guidance for its actual MPI runtime. Keep the runtime, transport setting, process placement and input workload together in the experiment record. Avoid carrying an old small-job setting into a larger run without examining that estimate.

\n

Verification

\n

Rehearse the candidate configuration with an approved representative job. Check correctness first, then compare completion behavior and measured scaling against the prior configuration. Change one intended variable at a time where practical so an improvement or regression can be attributed. Retain an unsuccessful result as useful evidence instead of turning the source’s tuning suggestion into a claimed performance guarantee.

\n

Official references

\n

Microsoft Learn: Scaling HPC applications. Source reviewed September 9, 2026.

", "content_text": "Source facts\nMicrosoft’s Azure HPC guidance estimates an MPI job’s maximum connections by multiplying processes per node by the square of the job’s node count. It suggests UCX_TLS=rc,sm for jobs below 256K connections and UCX_TLS=dc,sm above 256K connections. The source also emphasizes workload-specific tuning experiments rather than assuming an optimal configuration from VM selection alone. Microsoft Learn.\nApplicability\nUse this planning check when scaling an MPI job beyond an earlier test. Record both the process layout and node count; a node-count-only label does not express the source’s connection estimate. The documented smaller/larger examples do not specify the exact boundary case.\nDSE recommendation\nRecord the estimated connection count alongside the proposed MPI transport experiment. Have the application owner compare the intended job layout with the transport guidance for its actual MPI runtime. Keep the runtime, transport setting, process placement and input workload together in the experiment record. Avoid carrying an old small-job setting into a larger run without examining that estimate.\nVerification\nRehearse the candidate configuration with an approved representative job. Check correctness first, then compare completion behavior and measured scaling against the prior configuration. Change one intended variable at a time where practical so an improvement or regression can be attributed. Retain an unsuccessful result as useful evidence instead of turning the source’s tuning suggestion into a claimed performance guarantee.\nOfficial references\nMicrosoft Learn: Scaling HPC applications. Source reviewed September 9, 2026.", "content_markdown": "## Source facts\n\nMicrosoft’s Azure HPC guidance estimates an MPI job’s maximum connections by multiplying processes per node by the square of the job’s node count. It suggests UCX_TLS=rc,sm for jobs below 256K connections and UCX_TLS=dc,sm above 256K connections. The source also emphasizes workload-specific tuning experiments rather than assuming an optimal configuration from VM selection alone. [Microsoft Learn](https://learn.microsoft.com/en-us/azure/virtual-machines/compiling-scaling-applications).\n\n## Applicability\n\nUse this planning check when scaling an MPI job beyond an earlier test. Record both the process layout and node count; a node-count-only label does not express the source’s connection estimate. The documented smaller/larger examples do not specify the exact boundary case.\n\n## DSE recommendation\n\nRecord the estimated connection count alongside the proposed MPI transport experiment. Have the application owner compare the intended job layout with the transport guidance for its actual MPI runtime. Keep the runtime, transport setting, process placement and input workload together in the experiment record. Avoid carrying an old small-job setting into a larger run without examining that estimate.\n\n## Verification\n\nRehearse the candidate configuration with an approved representative job. Check correctness first, then compare completion behavior and measured scaling against the prior configuration. Change one intended variable at a time where practical so an improvement or regression can be attributed. Retain an unsuccessful result as useful evidence instead of turning the source’s tuning suggestion into a claimed performance guarantee.\n\n## Official references\n\n[Microsoft Learn: Scaling HPC applications](https://learn.microsoft.com/en-us/azure/virtual-machines/compiling-scaling-applications). Source reviewed September 9, 2026." }, "json_ld": { "@context": "https://schema.org", "@graph": [ { "@type": "Organization", "@id": "https://dsesecurity.com/#organization", "name": "Detection Systems & Engineering", "alternateName": "DSE Security", "url": "https://dsesecurity.com/", "logo": { "@type": "ImageObject", "url": "https://update.dsesecurity.com/assets/dse-logo-20260812.png?v=1.8.20" } }, { "@type": "Organization", "@id": "https://update.dsesecurity.com/#editorial-team", "name": "DSE Security Editorial Team", "url": "https://update.dsesecurity.com/", "parentOrganization": { "@id": "https://dsesecurity.com/#organization" } }, { "@type": "WebSite", "@id": "https://update.dsesecurity.com/#website", "name": "DSE Updates", "alternateName": "DSE Security Knowledge Hub", "url": "https://update.dsesecurity.com/", "inLanguage": "en-US", "publisher": { "@id": "https://dsesecurity.com/#organization" }, "potentialAction": { "@type": "SearchAction", "target": { "@type": "EntryPoint", "urlTemplate": "https://update.dsesecurity.com/?q={search_term_string}" }, "query-input": "required name=search_term_string" } }, { "@type": "WebPage", "@id": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/", "url": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/", "isPartOf": { "@id": "https://update.dsesecurity.com/#website" }, "lastReviewed": "2026-09-09" }, { "@type": "BreadcrumbList", "@id": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/#breadcrumbs", "itemListElement": [ { "@type": "ListItem", "position": 1, "name": "DSE Updates", "item": "https://update.dsesecurity.com/" }, { "@type": "ListItem", "position": 2, "name": "Size an MPI transport experiment by connection count, not node count alone", "item": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/" } ] }, { "@type": [ "Article", "TechArticle" ], "@id": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/#article", "identifier": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/", "url": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/", "headline": "Size an MPI transport experiment by connection count, not node count alone", "description": "What should be calculated before applying Azure HPC guidance for smaller versus larger MPI jobs?", "abstract": "What should be calculated before applying Azure HPC guidance for smaller versus larger MPI jobs?", "articleBody": "Source facts\nMicrosoft’s Azure HPC guidance estimates an MPI job’s maximum connections by multiplying processes per node by the square of the job’s node count. It suggests UCX_TLS=rc,sm for jobs below 256K connections and UCX_TLS=dc,sm above 256K connections. The source also emphasizes workload-specific tuning experiments rather than assuming an optimal configuration from VM selection alone. Microsoft Learn.\nApplicability\nUse this planning check when scaling an MPI job beyond an earlier test. Record both the process layout and node count; a node-count-only label does not express the source’s connection estimate. The documented smaller/larger examples do not specify the exact boundary case.\nDSE recommendation\nRecord the estimated connection count alongside the proposed MPI transport experiment. Have the application owner compare the intended job layout with the transport guidance for its actual MPI runtime. Keep the runtime, transport setting, process placement and input workload together in the experiment record. Avoid carrying an old small-job setting into a larger run without examining that estimate.\nVerification\nRehearse the candidate configuration with an approved representative job. Check correctness first, then compare completion behavior and measured scaling against the prior configuration. Change one intended variable at a time where practical so an improvement or regression can be attributed. Retain an unsuccessful result as useful evidence instead of turning the source’s tuning suggestion into a claimed performance guarantee.\nOfficial references\nMicrosoft Learn: Scaling HPC applications. Source reviewed September 9, 2026.", "datePublished": "2026-09-10T00:31:20+00:00", "dateModified": "2026-09-10T00:32:00+00:00", "mainEntityOfPage": { "@id": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/" }, "inLanguage": "en-US", "isAccessibleForFree": true, "author": { "@type": "Organization", "name": "DSE Security Editorial Team", "url": "https://update.dsesecurity.com/#editorial-team" }, "publisher": { "@id": "https://dsesecurity.com/#organization" }, "image": { "@type": "ImageObject", "@id": "https://update.dsesecurity.com/updates/dse-20260909-036-size-an-mpi-transport-experiment-by-connection-count-not-node-count-alone/#primaryimage", "url": "https://update.dsesecurity.com/assets/editorial/continuity-recovery-social-v2.jpg?v=1.8.20", "contentUrl": "https://update.dsesecurity.com/assets/editorial/continuity-recovery-social-v2.jpg?v=1.8.20", "width": 1200, "height": 630, "caption": "Size an MPI transport experiment by connection count, not node count alone" }, "articleSection": [ "Business Continuity", "IT" ], "keywords": [ "Business Continuity", "IT", "Guide", "Information priority" ], "genre": "Guide", "about": [ { "@type": "Thing", "name": "Business Continuity", "url": "https://update.dsesecurity.com/topic/business-continuity/" }, { "@type": "Thing", "name": "IT", "url": "https://update.dsesecurity.com/topic/it/" } ], "wordCount": 237, "timeRequired": "PT2M", "publishingPrinciples": "https://update.dsesecurity.com/updates/dse-updates-editorial-methodology/", "usageInfo": "https://update.dsesecurity.com/usage/", "copyrightHolder": { "@id": "https://dsesecurity.com/#organization" }, "copyrightNotice": "Copyright © 2026 Detection Systems & Engineering. All rights reserved.", "citation": { "@type": "CreativeWork", "name": "Scaling HPC applications - Azure Virtual Machines - Azure Virtual Machines | Microsoft Learn", "url": "https://learn.microsoft.com/en-us/azure/virtual-machines/compiling-scaling-applications" } } ] } }