Some thoughts on ADA deployment
D. Adams
13jun05

Clients
-------

Users interact through clients (dialroot, ganga, web, ...) and
submit abstract jobs (AJ's): transformation, dataset and preferences.

Services
--------

These requests are handled by a primary analysis service that examines
the request and forwards it to an appropriate local secondary service
based on the type of tranformation, data and preferences. Each tier 1
site hosts two primary analysis services:

PA1. To handle requests from users.

PA2. To handle requests from services running at other sites.

These might be combined.

Each tier 1 would run a number of secondary analysis services which
handle requests from the primary services at the site:

SA1. One to carry out local processing. Could be more than one for
     different performance requirements.

SA2. One for each site contributing analysis capabilities without
     supplying a analysis service. Submission to the site would be
     follow the local grid protocol.

SA3. Two to split requests between sites. One making use of this and
     other tier 1 sites and one using associated tier 2 and other sites.

SA4. One to handle large user production jobs using the ATLAS production
     system. Job database might include all of ATLAS or be restricted to
     the local tier.

Tier 2 sites are not required to deploy analysis services but can
provide analysis capablilites via an "analysis CE" that provides the
requisite performance (submission rate and latency, data throughput).
Or a tier 2 site may choose to deploy an SA1 service eliminating the
need for the tier 1 to run an SA2 service for that site.

Other sites including opportunistic non-ATLAS sites can contribute
in the same manner as tier 2 sites.

Requirements on participating sites
-----------------------------------

Sites contribute to ATLAS analysis by agreeing to take requests from
or more tier one sites and then providing an analysis service or
analysis CE to handle those requests.

The site must agree to provide storage for the files included in one
or more datasets and register those files in a local replica catalog.
The site must also provide temporary storage for output results
guaranteeing to keep the data for the lifetime assigned by the
assigning job or until it can be copied to the tier 1 site.

The site should provide some number (say at least 50) slots for
analysis job processing and guarantee that any analysis job submitted
when one or more slots is open will start in a timely manner. The
site should provide peak submission rates and control latencies so that
a collection of 100 trivial jobs can be completed in about a minute.

The site should provide sufficient throughput that its slots can all
be simultaneously used for I/O-intensive analysis jobs. At present,
ATLAS analysis runs at 2 MB/s implying a need for 400 MB/s total
throughput to handle 100 jobs.