Billing and usage

Set budget_seconds with this page in mind. The app shows the rule of each site, where there is one, and the seconds that the records hold. It does not claim what a site bills: the accounting of the site decides what your award was charged. What “QPU hours” mean at each site: the lesson page “QPU hours” on the devices.

Site	Rule of the site	What the records hold
VLQ	max(shots x 400 us, 0.9 s) per job, the worst case of the IT4I consumption guide; for a typical circuit a shot costs about 350 us plus the time of the circuit. Each job also has 3 to 15 s of latency, which IT4I does not bill. Init jobs book classical node time only	the device seconds the job reports, which can be below the 0.9 s floor; runs and usage show both
Euro-Q-Exa	EuroHPC time is billed in wall-clock QPU seconds, not in equivalent qubit-seconds, a unit LRZ also defines (LRZ)	the seconds from scheduled to completed, the app’s own measure
Piast-Q	no public rule	wall seconds, from the submission to the result, as Piast-Q gives no device time

The dry run prints the cost of each job by the published rule, and the rule of the site in its billing line.

The budget

Before each job the rule budget adds up:

• the measured seconds of every result in every record of the site and project, and in older_records (a negative time counts 0);

• job_seconds_limit for each order of unknown time: a lost or open order, a failed job, an order with no result that an older runner wrote as expired, and a result with no measured seconds;

• job_seconds_limit for the job about to be sent.

Above budget_seconds, the job is not sent and the run stops. Raise budget_seconds only if your award allows it. The sum is an upper bound of what the app has used, not the charge of the site.

• Measured seconds are the first number among seconds, device_seconds (older runners) and wall_seconds. wall_seconds (job request to result) is a number when the call that sent the job also collected it, else null. So a result with no seconds (for example a Euro-Q-Exa result whose time stamps cannot be read) counts its wall seconds, or else job_seconds_limit. jobs, runs and usage show them as measured, and resolve --over-limit judges on them. The rule job_time reads seconds only, so it stops at such a result.

• 0 seconds: a refusal, a job never submitted, and an order resolved with --none. But an order resolved as still New, or one whose job id no URL can hold, counts at job_seconds_limit: its job may yet run.

• After --over-limit, --expired or --gone: the counted_seconds of the resolved line, the job_seconds_limit of the config at the decision. A later change of the config does not change them. For a record of another runner, a result it writes later counts with its measured seconds (Resolving an order, and a stale lock).

• Other members: the app counts only the records it reads. LRZ counts the use of every member of a project against one budget (General information on the Euro-Q-Exa user operation). Set budget_seconds with the use of the other members in mind.

A result above the job time limit

The seconds the site reports can hold a wait at the site, cause not known. On Piast-Q, a result collected by a later call holds the wall seconds up to that call. The rule job_time then stops the run, and the result counts with all its seconds, so the rule budget can stop every later send of the project.

When you find that the time is such a wait, give qdevices resolve RECORD ORDER --over-limit. The result then counts at job_seconds_limit; its measured seconds stay in the record, and runs and usage show both. Decide it for each order: a job that did run that long would then count too little.

qdevices usage

qdevices usage [PATH ...] gives the totals by site and project (Commands). Seconds counted put the seconds of a decision of --over-limit, --expired or --gone in place of the measured ones. Neither is the sum of the budget: both leave out each order of unknown time, and an order resolved as still New. The totals come from the records found, not from the sites. To count older records, give their folders as PATH, or list them in older_records.