At a glance
This release makes the optimized execution path return the same complete result set as the standard path. That includes post-annotated execution traces and very large result sets. Simulation logs are now available for runs on that path, and three customer-reported issues are resolved. Simulator software moves to 4.7.1: the RoCE NIC now counts and reports packets it discards on receive. Chakra configurations that declare more trace files than the topology can place are now rejected before launch. The release also includes security updates. Three items reported by customers are resolved in this release. See Customer-reported issues resolved below.
Simulated behaviour is unchanged, but one report changes shape. The RoCE NIC now counts receive-path discards it previously dropped silently, and reports them in eight new roce-transport-stats columns. The new columns are inserted within the row rather than appended. Analysis that reads that report by column position must be updated; reading by column name is unaffected. See Simulator software 4.7.1.
Customer-reported issues resolved
Fetching simulation log files sometimes failed
Simulations that ran on the optimized execution path had no log to serve: the logs operation looked in a location those runs never wrote to, and returned “not found”. The logs operation now reads each simulation’s log from where that run stored it. The log carries the output of every stage of the run, including the simulator’s own. No change to the API. The same logs operation, parameters and response apply. Standard-path simulations are served exactly as before.Post-annotations were not available in execution traces
Chakra simulations on the optimized execution path did not return their post-annotated traces. They now deliver, for every rank in the traceset:- the post-annotated rank trace, and
- its rank-to-rank (R2R) file.
post-annotated/ in the simulation’s results, together with the cluster and compute-node totals. They are listed and downloaded through the existing results operations.
Simulations succeeded with empty results
On the optimized execution path, a simulation could report success even when its result files had not all been delivered. It then presented an empty or partial result set. Two changes close this:- A simulation now completes only when all required result files have been delivered. A delivery failure is reported as a failure, with the reason. It is no longer a success with missing output.
- Result delivery is built for very large runs. Uploads use bounded concurrency with per-file retries, sized for thousands of result files totalling on the order of 200 GB.
Complete results on the optimized execution path
On earlier releases, simulations served by the optimized execution path returned a subset of the files the standard path produces. With this release, the optimized path returns everything the standard path returns. The customer-reported fixes above follow from this. In addition:- Failed simulations keep their results. Every file delivered for a run can be listed and downloaded by its exact name, including runs that ended in failure.
- Failures during set-up report their real cause. If a simulation fails while it is being prepared, the original reason, exit code and failing step are preserved and reported. It is no longer collapsed into a generic failure.
- Results contain only simulation output. Internal run records and diagnostics are stored separately and never appear in a simulation’s result listing.
- Storage for the largest runs. Compute storage is now sized for dense 4,096-node launches, so the largest supported runs on this path do not exhaust disk while producing results.
New in this release
Simulator software 4.7.1
RoCE NIC receive-path discards are now counted and reported. The RoCE NIC discards received packets in two situations. Both are correct model behaviour:- when no queue-pair context exists for the packet’s flow;
- when a NAK is outstanding on the queue pair, in which case packets behind the gap are discarded until it fills.
roce-transport-stats, which gains a cumulative and an interval column for the total and for each reason:
No change to simulated behaviour. Only the transport report changes. See the note in At a glance about column positions.
Chakra configurations are checked against the topology
A Chakra application that declares more trace files than the topology has Chakra-capable hosts has no valid outcome. Previously this was reported only as a warning, and such a configuration could still be saved and launched. It is now rejected at validation with a specific message.A configuration that was accepted on 0.1.1427 may now be rejected. That is intentional: the configuration could not run correctly. The failure now arrives at validation with a clear reason, not during the run.
Model catalog changes
No model versions change in this release. The shipped default templates are updated to reference simulator software 4.7.1.Platform reliability and stability
- Launches that cannot reach their compute instance now fail promptly. If a launch could not reach or authenticate to its compute instance, the simulation could previously remain pending indefinitely while the instance stayed provisioned. It now ends with a dispatch-error failure, and the request can be resubmitted.
Security updates
This release contains security updates. We recommend upgrading. In keeping with our disclosure practice, we do not publish details of security fixes. Customers with a specific compliance question should contact their Scala representative.Documentation
The simulation output-file reference is updated for the newroce-transport-stats columns.
Not included in this release
The following remain open and are not addressed by 0.1.1828:- Summary metrics for very large result sets. For multi-gigabyte result sets, the simulation summary view may not return metrics. Raw result files remain available through the results download operations.
- Uploading custom Chakra mapping files. Related operations may appear in the API specification but are not yet enabled.
- Pre-simulation run-time estimation. There is currently no supported way to determine in advance whether a trace will exceed the configured simulation time limit.
- ECN-marked packet counts received at the NIC. NIC-side receive counters for ECN-marked packets are not yet reported. Switch-side ECN information remains available.
Upgrade notes
- Simulator software 4.7.1 is applied automatically with the platform upgrade. No configuration change is required.
- Existing configurations continue to work and behave identically, except Chakra configurations that declare more trace files than the topology can place; these are now rejected at validation (see above).
- Update positional readers of
roce-transport-stats. If you parse that report by column position, switch to column names or account for the eight inserted columns.