|>Top >Feature Story >this page||
last update: 10/05/12
|Designing the ultrafast
DAQ for Belle II
KEK’s Belle experiment has played an important role in particle physics
for more than a decade. Researchers are currently hard at work on a
planned upgrade. The new detector, the Belle II, will be hundreds of
times more powerful than the original. One component which is central
to the Belle II detector is the data acquisition system, DAQ. Learn here
what four experts at KEK are planning to beat the many gigabytes of
data estimated to hit their system every second.
The DAQ for Belle II will be extremely complex, owing to the sheer size of data and the frequency of events. “The detectors will trigger 20 thousand times each second, and each trigger produces a signal that is 300 kilobytes. This means that the DAQ must deal with 6 gigabytes of data every second,” explains Prof. Ryosuke Itoh of KEK, the head of the DAQ team for both Belle and Belle II. That data rate is an increase of a factor of about 40 from Belle. “Because we cannot store everything we receive, the DAQ implements a series of processes to cut down the number of events to only those which are physically interesting.”
The Belle II team’s new DAQ scheme will be based on three ideas: smooth transition from Belle to Belle II, unification of subsystems, and scalability. During the upgrade, the DAQ will undergo changes in almost all its components, although the original Belle DAQ design philosophy will remain intact.
The main components of the DAQ are: the front-end digitizer, the unified data link (Belle2Link), the common readout platform (COPPER), the event builder, and the high level trigger (HLT). One noticeable change from the original Belle DAQ is that the digitization will be done by the electronics installed near the detectors, instead of on the COPPER board. Each front-end digitization module will handle 100 detector channels. There are millions of channels in total. The digitized data produced by the individual detectors will be merged and then transmitted through the Belle2Link to the COPPERs. The role of the COPPERs is to receive the merged data and place it on the network so that readout PCs can bring data from all modules together to prepare for event building. Based on the reconstructed physical information of each event, the HLT software then decides if the event is to be stored.
Standardizing the data structure
The entire process where the DAQ controls the data, from the digitization at the individual detectors until storage, is called the ‘online’ process. Any data handling after the data are stored is called an ‘offline’ process. The most important, unique feature of the Belle II’s DAQ system is that the software used for both online and offline for data handling is built on a common framework called basf2.
One important change in the software is that Itoh and the offline team have written the DAQ software framework and the offline software components in an object-oriented language, to adopt to more sophisticated data handling scheme. Itoh is currently working on the implementation of basf2 in HLT, COPPER, and readout PCs.
Itoh is responsible for building the HLT as well. “The HLT must process a large amount of data in short period of time,” says Itoh. “Because each event is isolated, multiple events can be processed simultaneously.” Itoh, together with student Soohyung Lee of Korea University, is now working on implementing parallel processing to increase the number of processing nodes.
In the HLT, there are two levels of parallel processing. The first level enables a computer with a multi-core processor to distribute events among the cores within that computer to process events in parallel. The second level enables the distribution of events among multiple computers each with a multi-core processor. The team has already completed the first level of parallel processing, and is now working on the second level. The key, Itoh says, is careful data handing from computer to computer.
The low-maintenance, simple event builder
Prior to the data storage, the HLT needs to judge whether the event is physically interesting and worth storing. The event to be stored also needs to include a ‘complete’ set of data associated with the event. For this, the event builder system packs data fragments that belong to the same trigger into an event data. Prof. Soh Suzuki of KEK, responsible for building the event builder, stresses the importance of having a low-maintenance system and using simple coding practices.
Suzuki plans to use a network switch to control the data flow between PCs for event building. The usual difficulty with the network switch is to handle the size of the fully built event packets. At Belle II, each event packet may be as large as 300 kilobytes, and no network switch is commercially available to store such a large amount of data in the buffer to prevent packet loss due to data overflow at output ports. Suzuki utilizes a technique called the 'barrel shifter'. The barrel shifter rotates connections among a series of computers in a cyclic sequence, switching from node to node as necessary. In total, the system requires around 50 inputs for the readout PCs and around 10 outputs for the HLT units.
The key feature in Suzuki's design is that he arranges PCs in mesh to produce required numbers of input and output ports. First, he uses four 4-input, 4-output barrel shifters in two rows to create a 16-input, 16-output barrel shifter. In the next step, he connects four of 16-input, 16-output barrel shifters with sixteen of 4-input, 4-output barrel shifters, to create a 64-input, 64-output barrel shifter. This is now large enough to host the required number of input and output ports. With this design, the number of required PCs is reduced to 48. A prototype of this system was developed last year, proving the idea to be feasible.
Suzuki is also working on the event builder software that efficiently handles the high data rate. The Belle II event builder will receive data packets of different size at different timing, and will still have to process six gigabytes every second. At the time of Belle, PCs did not have sufficient processing power to handle the entire data stream in one process, and so had to utilize a multi-process system in which processes were closely tied via inter-process communication techniques, such as shared memory, semaphore, and message queue. The many processes made it difficult to understand the system behavior, especially when something went wrong. It was the recent evolution of the Linux operating system and the dramatic increase of the CPU power that made it possible to build a single process event builder unit. Suzuki built a prototype and examined its performance under the expected data rate of Belle II. "The result shows that the latest generation CPU-based PC has a sufficient power for a 4-input, 4-output barrel shifter unit," says Suzuki.
The DAQ team already completed the COmmon Pipeline Platform for Electronics Readout, COPPER, five years ago. COPPER is a general purpose pipeline readout board, whose aim is to reduce dead time. Dead time is the time after each data readout during which the DAQ cannot handle another incoming data.
As a live test, scientists replaced most of the previous system at Belle with COPPER. The system has run successfully ever since. In the Belle system, COPPER has four daughter cards, collectively called FINESSE (Front-end INstrumentation Entity for Sub-detector Specific Electronics). These cards take care of data digitalization. The data then is transferred to a CPU on COPPER. The CPU is like a tiny computer, on which relatively complex procedures can be easily coded. As a result of the COPPER installation, the dead time of the central drift chamber was reduced by 90%.
For the Belle II, the digitizers on FINESSE will be moved to the front-end electronics installed inside or near each detector. In this new design, FINESSE will act as a receiver of data sent through optical fiber from the detector electronics. “If the digitizers remain on COPPER, the number of cables required to transmit the analog signals to FINESSE would be enormous,” explains Dr. Takeo Higuchi who’s been in charge of designing, developing, and implementing COPPER ever since the inception of the COPPER project eight years ago. “For Belle II, each detector has freedom to decide on an optimal digitization scheme.”
The most challenging detector for DAQ is the newest detector component, the inner most detector—the pixel detector (PXD). Since the PXD is the closest detector to the collision point, the size of data produced by each event is enormous—1 megabyte. “This is beyond the COPPER’s capability,” says Higuchi. There are currently three options proposed by groups from Germany and KEK. The data from PXD is readout either by a new hardware or a software component, and handled by a new software component. Higuchi coordinates the collaborative efforts between them. “Germany’s hardware solution will be an innovative technology, and may be a big challenge.” The team is now proposing detailed plans that would meet their deadline.
A unified data link
Belle II will have to handle the dramatically increased number of detector channels and amount of data. This demands that the digitization be done already inside or just near the detector. The digitized data produced by these "front-end" digitizers will be sent through a 30-meter long unified data link, Belle2Link. Belle2Link packs and serializes many channels of digitized data into a single optical fiber line. Another DAQ expert Prof. Mikihiko Nakao of KEK, together with the group led by Prof. Zhen-An Liu of Institute of High Energy Physics (IHEP) in China, is working on this data transmission from point A to point B. Building the link, however, is much harder than it may sound, as this is the place where the high trigger rate of the Belle II creates huge amount of data.
The main concept of the Belle2Link is again unification. It unifies the hardware design, firmware, and protocols. The interfaces to the detectors are also designed to handle different types of data in one common framework. “If the readout logic varies detector by detector, it is harder for DAQ to handle the data, so the logic is integrated and all detector electronics use the same logic,” says Nakao. Owing to this, the DAQ will be able to talk to the detector boards directly.
Nakao's main responsibility is the distribution of the trigger signal to the front-end digitizers. In order to guarantee the synchronization of the entire Belle II DAQ system, “we need many more signals on top of the trigger signal itself. More signals mean more cables for transmission. By serializing these related signals into one signal, we can reduce the number of cables required for transmission,” explains Nakao. He is now working on design and development of the circuit boards, and the firmware that will drive the boards.
The gang of four
For the large amount of demanding work necessary for the DAQ, the team is relatively small. Four researchers at KEK, in collaboration with international teams, are going to rebuild the entire DAQ system for Belle II. The core members are the four who have built, maintained, and continually upgraded the Belle DAQ for the past two decades.
“Everyone here is very capable, and things are moving swiftly,” says Itoh. “We are aiming for a stable, low-maintenance system design. Also, we will need to scale up our system as Belle II’s luminosity advances. The Belle DAQ was an evolving technology, and the Belle II DAQ will also be an evolving technology.” Perfecting this system will require the full abilities of all four experts.
|copyright (c) 2010, HIGH ENERGY ACCELERATOR RESEARCH ORGANIZATION, KEK
1-1 Oho, Tsukuba, Ibaraki 305-0801 Japan