Skip to main content

Cornelis Technical Documentation

4.4.2. Parameters and Impacts on the System

Parameters that may disrupt a live system should not be changed while applications are running. To minimize disruptions, Cornelis recommends the following steps:

  1. Stop all Fabric Managers.

  2. Move all host-to-switch links to Init, which stops all fabric applications, using either of these methods:

    • Reset all hosts or bounce their fabric links.

    • Reset all Switches.

  3. Complete the change to the Fabric Manager configuration.

  4. Restart all the Fabric Managers.

Changes to the following parameters may disrupt some applications or impact application performance and must be carefully considered on a live fabric:

  • Changing PathRecord Settings

    Applications that use PathRecords (PathSelection) to select end-to-end addresses may ignore any resultant changes and continue to use the previous settings until the application is restarted or it re-fetches its PathRecords. This can change what paths applications use (only applicable when LMC != 0) and cause application performance changes or impacts.

  • Activating or Deactivating Pre-Defined, Pre-Enabled vFabrics

    This is intended to be permitted on a live fabric; however, applications attempting to use a deactivated vFabric will fail after the vFabric is deactivated. In general, such applications should be stopped prior to deactivating the given vFabric. Activation of a pre-defined pre-enabled vFabric is typically safe.

  • Changing Device Groups or Applications Associated with an Active vFabric

    Applications or nodes whose assignments have changed may fail or continue executing with the old PathRecords and therefore not fully obey the change.

  • Changing an Active vFabric's Bandwidth, Timeouts, Priority, or Preemption/Traffic Flow Optimization Rank

    This will change arbitration and scheduling within the fabric. Application operation should not be impacted; however, the performance and scheduling of fabric bandwidth will change to the new settings in a timely manner. Applications that use PathRecords to compute end-to-end timeouts may ignore any resultant changes in their timeouts and continue to use the previous timeouts until the application is restarted or refetches its PathRecords. Typically, changes to timeouts are small and have limited impact, but under extreme congestion situations, this could cause unexpected application performance impacts.

  • Changing Fabric Manager Virtual Lane (VL) Buffer Allocation Parameters

    These parameters tune VL buffering and therefore affect application performance:

    • MinSharedVLMem

    • DedicatedVLMemMulti

    • WireDepthOverride

    • ReplayDepthOverride

  • Changing Fabric Manager VL Buffer Allocation Parameters

    These parameters tune VL buffering and therefore affect application performance.

  • Changing Parameters That are Activated Only When a Port Is Bounced

  • These parameters tune preemption/traffic flow optimization and therefore affect application performance

  • Changing CongestionControl:

    • CongestionControl

  • Changing Routing Algorithm:

    These can result in rerouting of the fabric and may briefly disrupt traffic. If incorrect (or better than previous) choices are made, fabric and application performance can be greatly affected (for better or worse).

    • RoutingAlgorithm

    • SpineFirstRouting

    • ForceRebalance

    • FatTreeTopology

    • DGShortestPathTopology

    • AdaptiveRouting

  • Using ForceAttributeRewrite

    This can undo previous adaptive routing decisions in switches and result in a short-term change to application performance.

  • Changing Fabric Manager Security Parameters

    Accurate changes typically do not affect existing applications. However, security fixes or stricter policies may disable applications or nodes that previously operated in the cluster. This may be intentional when the goal is to strengthen security.

    • PreDefinedTopology

    • VL15CreditRate

    • SmaSpoofingCheck

    • McDosThreshold

    • McDosInterval

    • McDosAction

    • SmAppliance

  • Changing Fabric Manager Multicast

    Changes to FM Multicast section of parameters may impact assignments of Multicast Local Identifiers (MLIDs) to multicast groups or QoS parameters (Mtu, Rate) or Security parameters (PKey) associated with a given multicast group. Typically on FM restart, end nodes must rejoin multicast groups, and if these parameters change there can be minor disruptions of multicast traffic via lost packets. Most multicast applications will ride through these short-term disruptions without error.

  • Changing Fabric Timeouts

    On uncongested fabrics, minor changes often do not affect applications. Significant decreases in values may cause congestion mitigation mechanisms in switches to fire sooner, resulting in packet discards and changes to application performance. Significant increases may slow switch mitigation and allow congestion to propagate further and impact application performance.

    • TimerScalingEnable

    • SwitchLifetime

    • HoqLife

    • VLStallCount

  • Changing Link Policies

    If all active links are currently "in policy" this will have no impact on existing applications. However, mistakes or tightening of link policies may disable nodes or links in the cluster that were running prior to the change. Such changes may be intentional and desired if the goal is to tighten link policies and impact such nodes or links.

    • HFILinkPolicy

    • ISLLinkPolicy

  • Changing Subnet Administrator (SA) Timeouts

    This change affects SA query timeouts and request handling under heavy load. Most applications see little impact, but application behavior and performance may change during extreme conditions.

    • SaRespTime

    • NoReplyIfBusy

  • Changing FM SubnetSize

    This alters the memory used by the FM for selected buffering and may impact FM performance. In general, there should be no impact on existing applications beyond changes to the FM's responsiveness to Performance Manager (PM) and Performance Administrator (PA) requests.

  • Changing QueryValidation

    Cornelis recommends not changing this, as non-compliant OpenFabrics Alliance (OFA) applications may cease to work if it changes from 0 to 1.

  • Changing LID

    This may change the LID of the Subnet Manager (SM) and briefly disrupt application queries to the SA.

    Parameters that can safely be changed on a live system without impacting applications are the following:

    • PM/PA parameters (may impact active management and monitoring applications, but will not impact non-management applications running on the cluster).

      This includes adding or removing PM port groups, altering DeviceGroups that are only used as PmPortGroups.

    • FM controls on FM core dumps (CoreDumpLimit, CoreDumpDir)

    • Changes to FM failover and DB Sync parameters (Priority, ElevatedPriority, ConfigConsistencyCheckLevel, MasterPingInterval, MasterPingMaxFail, DbSyncInterval)

      However, such changes may result in a different FM becoming primary or the movement of inconsistent FMs to standby.

    • Changes to FM sweep time, Subnet Management Agent (SMA) timeouts/retries, SM strategies (VL15FlowControlDisable, SmaBatchSize, MaxParallelReqs, SweepInterval, IgnoreTraps, MaxAttempts, RespTimeout, MinRespTimeout, SweepErrorsThreshold, SweepAbandonThreshold, TrapThreshold, TrapThresholdMinCount, NonRespTimeout, NonRespMaxCount, SwitchCascadeActivateEnable, NeighborNormalRetries)

      This mainly impacts SM responsiveness to fabric changes and SM handling of non-responsive nodes. Typically has no application impact. However, if adjustments are such that nodes are now not responsive within expected time frames, SM could drop some nodes from the fabric.

    • Changes to FM logging (LogLevel, LogMode, SyslogFacility, Debug, RmppDebug,*LogMask*, NodeAppearanceMsg, SmPerfDebug, SaPerfDebug, PortBounceLogLimit)