EVM Project Status

1
EVM Project Status

• People
• S. Aziz
• M. Litmaath
• C. Moore
• V. ODell
• S. Pavlon
• K. Sumorok
• I. Suzuki
• Institutes
• FNAL
• MIT

2
Contents

• Test bench hardware
• IEEE1394 benchmark test
• BM software in XDAQ
• Simulation
• Plans

3
EVM Test Bench Updates

• Four additional nodes (18 in total)
• Myrinet (PCI64A), 100 Mbps Ethernet (public
  private), IEEE1394
• VME PowerPC board MVME2300
• RM, IEEE1394 bus fan-out board development
• S. Aziz (50 ) joined the project

UP
PPC
FastEthernet
IEEE1394
SMP
Myrinet
4
IEEE1394 Benchmark (S.Pavlon)

• Linux 2.2.14 kernel beta driver 20000616
• D-Link IEEE1394 PCI I/F (400 Mbps)
• Asynchronous write
• Latency 53 µs
• Bandwidth 34 MB/s

5
EVM Software on XDAQ

• XDAQ BM modified
• stores EventID in object shared by BM and RM
• EventID lifetime
• number of allocated EventIDs
• etc.
• keeps statistical info in ROOT
• histograms and n-tuples
• BM worked on FNAL test bench and CERN machines
• RM code under development
• original XDAQ RM was used in tests

6
XDAQ Benchmark at FNAL

• Configuration
• One unit (EVM/RU/BU) driven by one XDAQ executive
  on each machine
• Block size 4 kB, request 20, 10 msg/packet
• Event throughput
• 1 to 1 0.5 kHz _at_ EVM, 1.85 MB/s _at_ BU
• 1 to 2 1.0 1.85
• 1 to 3 1.8 1.85
• 2 to 1 0.5 3.72
• 2 to 2 1.0 3.72
• 3 to 1 0.5 5.58

7
XDAQ Benchmark at FNAL (cont.)

• 1 to 1 20 events active in the system
• Event throughput proportional to BUs
• Transfer rate behaves strangely
• Bandwidth sometimes twice higher in first 10
  seconds
• 1 to 3 BUs do not receive events evenly
• Further study needed
• Statistical information in the BU may help

8
EVM Simulation

• EVBSIM new components
• BCN using BCNICs
• BDN using BDNICs
• NICs ensure each endpoint does only one
  send/receive at a time
• currently BCNIC can send and receive at the same
  time (easy to change)
• Problems
• 500x500 cannot keep up with trigger rate
• RU does not implement barrel shifter not trivial
• BU traffic shaping may be the solution
• 500x500 memory usage gt 2 GB !

9
EVBSIM Modules
The given numbers are typical for LHC operation.
10
NxN Test Stand Setup

• 100 Mbps single private Ethernet
• Latency 72 us, BW 9.6 MB/s TCP
• Used for BCN and BDN
• EVBSIM uses 2 networks
• Not a problem here
• 1x1 and 2x2 test cases
• 0.0 ms L2, using all RUs
• Trigger rates
• Observed in XDAQ
• Input for EVBSIM
• 0.5 kHz 1x1, 1.0 kHz 2x2
• 1 BCN packet of 160 B per 10 ev.
• 1 BDN packet of 2 kB per 1 event
• Expected latencies
• 3 (3 1 in parallel) BCN and 10 (20) BDN
  messages per batch
• 1x1 3 (72 160/9.6) 10 (72 2048/9.6)
  3119 us
• 2x2 3 (72 160/9.6) 20 (72 2048/9.6)
  5973 us

11
NxN Test Stand Comparison

• XDAQ

EventID
1 to 1
2 to 2
lifetime
EVBSIM
12
EVM Plans

• EVM SW development on XDAQ
• RCN SW development benchmark
• IEEE1394 evaluation
• Understand test stand latencies
• Discuss how to proceed with EVBSIM
• Implement BU traffic shaping
• Compare with test stand anyway
• Run control for XDAQ (C. Moore)