1. Dr Andrew Peter Hammersley ESRF ESRF MX COMPUTIONAL AND NETWORK CAPABILITIES (Note: This is not my field, so detailed questions will have to be relayed to the relevant people. Alexander Popov of the ESRF MX Group is a remote participant.)

2. CURRENT ESRF MX PROCESSING HARDWARE ESRF has 6 (or 7) MX End-stations (all can be used simultaneously) Eiger detectors (1 for MX and 1 for Soft Matter) connected by 10 Gbit Ethernet to “Local Buffer Storage” and on to storage disks and a dedicated MX processing cluster General Parallel File System (GPFS) is used for high throughput to / from disk MX Processing Cluster: Linux 64, Debian 7 mxproc BULLX 9 blades, 108 cores, X5675 3.06GHz 6-core 147Gflops, 2GB RAM per core mxnice Dell C8220, 9 blades 180 cores, E5-2680v2 2.8GHz 10-core 227Gflops, 3.2GB 1600MHz RAM per core

3. ESRF PROCESSING CLUSTER Courtesy of E. Eyer, ESRF

4. Systems and Communications Infiniband To X8 CR106 To X8 CTB054 Dasnet 7 L3 Public network 232 Router (X8) Old Detector PCs without 1Gbe/10Gbe 202 Ge1net 202 Old Detector PCs with 1Gbe private (ge1) 198 (NFS) Compute Clusters (all have a 10ge1net interface) 197 (NFS) Slow Detector PC with 10Ge SFEprivate SFE Dectris Eiger-4M PC 197 (GPFS) GPFS admin 233 233 (GPFS) Detector PC With GPFS client 196 10ge1net 196/197/198 197 (NFS) LBSv1 Detector PC with 10Gbe and LBSv1 LBSprivate SFA HDDs Directmon NSD SFE private SFE Fast Detector PC with 10Gbe gz X 2 X 2+ ESRF NETWORK SCHEMATIC NFS CIFS NFS CIFS NSD

5. CURRENT ESRF MX PROCESSING SOFTWARE User Interface / Control Software: MXCube Viewing and Quick-Look Quality Control: ADXVIEW, ALBULA for viewing; DOZOR (Alexander Popov) for B-factor. ALSO running on general MX processing cluster Integration Software: Mainly XDS (MOSFLM used for screening crystals) ~80 – 90% of users use XDS at present Typically 3 processing pipe-lines are used Eiger File Conversion: The Eiger HDF5 files must be converted to be input into XDS or MOSFLM. Conversion programs are called as sub-processes to avoid creating extra files. NOTE: The two converter programs are NOT interchangeable! Diagnostics: A test program is run which should take 1 minute at low load.

6. CURRENT ESRF MX PROCESSING STATUS Current Processing status: ~80% of the time, the one minute program runs in roughly one minute However in ~20% of the time it can take up to ~10 times longer This is unacceptable, and can lead users to think that the programs have crashed. It is thought that CPU overload is the main problem Solution: A hardware upgrade is in process

7. ESRF MX DATA PROCESSING UPGRADE Add 8 new blades: 224 cores, E5-2680v4 2.4GHz 14-core, 2.3GB 2400MHz RAM per core Processing capacity should be doubled (DIALS should be added as another processing pipe-line) The upgraded cluster should be available for MX processing in September 2016 The “Local Buffer Store” is no longer necessary as GPFS is performant enough direct to disk. This is due to phased out in ~1 year. A “Local Preview Processor” may be added instead to provide dedicated processing power for preview / data quality analysis

8. ESRF: PRESENT HIGH RATE DATA MX EXPERIMENT Up to 750 images per second Bursts of 1000 – 2000 images Typical data collection: 180 degrees in 0.1 degree slices; 100 images per second with 18 seconds collection, ~3 Mbytes per image ~300 Mbytes per second Exceeds 1Gbit Ethernet Time between experiments: Min ~10 seconds / Typical 5 minutes

9. Current MX Eiger Set-up Data Reduction Problem 100 Tbytes Experiment; Reduce (azimuthal integration) by factor 10 Investigating “Remote Direct Memory Access” for fast transparent data access. The data reduction processor will access the detector data automatically. 35 Gbits per second measured Proof of method by End of June Concept could be used for MX for previewing and quality control ESRF CURRENT DATA RATE CHALLENGE Detector PC GPFSGPFS 10 Gbit Dedicated Link Local Buffer Store Infiniband MX Processing Cluster Detector PC GPFSGPFS 40 Gbits Dedicated Link SSD’s GPU’s Reduction Processor Infiniband Processing / Storage RDMA

10. ESRF: FUTURE CHALLENGES Project for XFEL type data collection Embed crystals in gel with “0.5 crystals per shot on average” Squeeze gel continuously through beam Collect data at 500 Hz (750 Hz works, but not continuously for long periods of time) 1.8 million images per hour. Roughly 6 Tbytes per hour Each crystal randomly orientated Data set 20,000 to 30,000 diffraction images, from ~100,000 images Integration with CrystFEL and/or cctbx Where will the processing power come from? (Initially one-off experiments) (Big ESRF wide problem: Too many files for the file systems!)

11. ACKNOWLEDGEMENTS THANKS FOR PROVIDING THE INFORMATION: Olof Svensson: Data Analysis Unit David Von Stetten: MX Group Jens Meyer: Beam-line Control Unit Bruno Lebayle: Systems and Communications Andy Götz: Software Group