ATLAS RoI Builder + CDF ● Brief reminder of ATLAS Level 2 ● Opportunities for CDF (what fits - what doesn't) ● Timescales (more of what fits and what doesn't)
Conclusions from quick tour of ATLAS TDAQ ● ATLAS Level 1 looks very much like CDF Level 2 – hardware based clustering, energy sums, isolation and muon Pt ● Level 1 accept rate for ATLAS is similar to that for CDF (~ kHz) ● RoI Builder (RoIB) plus Supervisor processors could act in the same capacity as the alphas do in CDF – same transaction rate (L1A rate) – provided the decision code is simple Supervisors could make the full level 2 decision instead of I/O to LVL2 farm
CDF != ATLAS ● Input to RoIB is via S-link – this is different from CDF Run IIa but not different in IIb ● Latency in ATLAS Level 2 is enormous – buffers are ~1000 events deep not 4 – RoIB latency is a non-issue for ATLAS ● initial prototype of RoIB waited for all event fragments to arrive before sending - this alone would introduce unacceptable latency for CDF it also limits size of L1 RoI data (since this needs to be buffered in the RoIB) – Level 2 decision order is different from the L1A order
Potential CDF Level 2 ● Single Builder and input card ● 4 processsors – could be more ● Output of trigger decisions to PULSAR is probably reasonable, but not necessary
Beginning to look at simulation
Still too infantile but... 40kHz livetime 86% Latency in microseconds
Current Architecture 4k x 36b FIFO S-Link Connector GbE MAC LSI L8101 GbE SerDes Agilent HDMP-1636A Opt. Conn. Features – Large input FIFO – Large Buffer – Optical GbEthernet Physical layer – Conforms to S-Link specs – FPGA can do data alteration if needed 512k x 32b Memory FPGA
ATLAS plan/timescale ● The current plan is to produce a prototype early this year (~April) ● There is already a 12U prototype (has been for several years) which demonstrated 100kHz operation ● 9U prototype system will include the ability (important for CDF but not necessary for ATLAS) of “routing” as opposed to “store and forward” type operation
Conclusions ● Appears reasonable to think of RoIB as merging unit for CDF – Still work to do in demonstrating that this idea actually makes sense (same work needed for any other merging scheme) ● Next steps should work out dataflow (how much from which sources and including as many of the overheads realistically as possible)