I am sure everybody else has 10gige and 40gige networks and are sending terabytes of data before breakfast.

Myself, I only have one computer with a 10gige network link and sufficient number of daq boards to fill
it with data. Here is my success story of getting all this data through MIDAS.

This is the anti-matter experiment ALPHA-g now under final assembly at CERN. The main particle detector is a long but 
thin cylindrical TPC. It surrounds the magnetic bottle (particle trap) where we make and study anti-hydrogen. There are 
64 daq boards to read the TPC cathode pads and 8 daq boards to read the anode wires and to form the trigger. Each daq 
board can produce data at 80-90 Mbytes/sec (1gige links). Data is sent as UDP packets (no jumbo frames). Altera FPGA 
firmware was done here at TRIUMF by Bryerton Shaw, Chris Pearson, Yair Lynn and myself.

Network interconnect is a 96-port Juniper switch with a 10gige uplink to the main daq computer (quad core Intel(R) 
Xeon(R) CPU E3-1245 v6 @ 3.70GHz, 64 GBytes of DDR4 memory).

MIDAS data path is: UDP packet receiver frontend -> event builder -> mlogger -> disk -> lazylogger -> CERN EOS cloud 
storage.

First chore was to get all the UDP packets into the main computer. "U" in UDP stands for "unreliable", and at first, UDP 
packets have been disappearing pretty much anywhere they could. To fix this, in order:

- reading from the udp socket must be done in a dedicated thread (in the midas context, pauses to write statistics or 
check alarms result in lost udp packets)
- udp socket buffer has to be very big
- maximum queue sizes must be enabled in the 10gige NIC
- ethernet flow control must be enabled on the 10gige link
- ethernet flow control must be enabled in the switch (to my surprise many switches do not have working end-to-end 
ethernet flow control and lose UDP packets, ask me about this. our big juniper switch balked at first, but I got it 
working eventually).
- ethernet flow control must be enabled on the 1gige links to each daq module
- ethernet flow control must be enabled in the FPGA firmware (it's a checkbox in qsys)
- FPGA firmware internally must have working back pressure and flow control (avalon and axi buses)
- ideally, this back-pressure should feed back to the trigger. ALPHA-g does not have this (it does not need it).

Next chore was to multithread the UDP receiver frontend and to multithread the event builder. Stock single-threaded 
programs quickly max out with 100% CPU use and reach nowhere near 10gige data speeds.

Naive multithreading, with two threads, reader (read UDP packet, lock a mutex, put it into a deque, unlock, repeat) and 
sender (lock a mutex, get a packet from deque, unlock, bm_send_event(), repeat) spends all it's time locking and 
unlocking the mutex and goes nowhere fast (with 1500 byte packets, about 600 kHz of lock/unlock at 10gige speed).

So one has to do everything in batches: reader thread: accumulate 1000 udp packets in an std::vector, lock the mutex, 
dump this batch into a deque, unlock, repeat; sender thread: lock mutex, get 1000 packets from the deque, unlock, stuff 
the 1000 packets into 1 midas event, bm_send_event(), repeat.

It takes me 5 of these multithreaded udp reader frontends to keep up with a 10gige link without dropping any UDP packets. 
My first implementation chewed up 500% CPU, that's all of it, there is only 4 CPU cores available, leaving nothing
for the event builder (and mlogger, and ...)

I had to:
a) switch from plain socket read() to socket recvmmsg() - 100000 udp packets per syscall vs 1 packet per syscall, and
b) switch from plain bm_send_event() to bm_send_event_sg() - using a scatter-gather list to avoid a memcpy() of each udp 
packet into one big midas event.

Next is the event builder.

The event builder needs to read data from the 5 midas event buffers (one buffer per udp reader frontend, each midas event 
contains 1000 udp packets as indovidual data banks), examine trigger timestamps inside each udp packet, collect udp 
packets with matching timestamps into a physics event, bm_send_event() it to the SYSTEM buffer. rinse and repeat.

Initial single threaded implementation maxed out at about 100-200 Mbytes/sec with 100% busy CPU.

After trying several threading schemes, the final implementation has these threads:
- 5 threads to read the 5 event buffers, these threads also examine the udp packets, extract timestamps, etc
- 1 thread to sort udp packets by timestamp and to collect them into physics events
- 1 thread to bm_send_event() physics events to the SYSTEM buffer
- main thread and rpc handler thread (tmfe frontend)

(Again, to reduce lock contention, all data is passed between threads in large batches)

This got me up to about 800 Mbytes/sec. To get more, I had to switch the event builder from old plain bm_send_event() to 
the scatter-gather bm_send_event_sg(), *and* I had to reduce CPU use by other programs, see steps (a) and (b) above.

So, at the end, success, full 10gige data rate from daq boards to the MIDAS SYSTEM buffer.

(But wait, what about the mlogger? In this experiment, we do not have a disk storage array to sink this
much data. But it is an already-solved problem. On the data storage machines I built for GRIFFIN - 8 SATA NAS HDDs using 
raidz2 ZFS - the stock MIDAS mlogger can easily sink 1000 Mbytes/sec from SYSTEM buffer to disk).

Lessons learned:

- do not use UDP. dealing with packet loss will cost you a fortune in headache medicines and hair restorations.
- use jumbo frames. difference in per-packet overhead between 1500 byte and 9000 byte packets is almost a factor of 10.
- everything has to be done in bulk to reduce per-packet overheads. recvmmsg(), batched queue push/pop, etc
- avoid memory allocations (I has a per-packet std::string, replaced it with char[5])
- avoid memcpy(), use writev(), bm_send_event_sg() & co

K.O.

P.S. Let's counting the number of data copies in this system:

x udp reader frontend:
- ethernet NIC DMA into linux network buffers
- recvmmsg() memcpy() from linux network buffer to my memory
- bm_send_event_sg() memcpy() from my memory to the MIDAS shared memory event buffer

x event builder:
- bm_receive_event() memcpy() from MIDAS shared memory event buffer to my event buffer
- my memcpy() from my event buffer to my per-udp-packet buffers
- bm_send_event_sg() memcpy() from my per-udp-packet buffers to the MIDAS shared memory event buffer (SYSTEM)

x mlogger:
- bm_receive_event() memcpy() from MIDAS SYSTEM buffer
- memcpy() in the LZ4 data compressor
- write() syscall memcpy() to linux system disk buffer
- SATA interface DMA from linux system disk buffer to disk.

Would a monolithic massively multithreaded daq application be more efficient?
("udp receiver + event builder + logger"). Yes, about 4 memcpy() out of about 10 will go away.

Would I be able to write such a monolithic daq application?

I think not. Already, at 10gige data rates, for all practical purposes, it is impossible
to debug most problems, especially subtle trouble in multithreading (race conditions)
and in memory allocations. At best, I can sprinkle assert()s and look at core dumps.

So the good old divide-and-conquer approach is still required, MIDAS still rules.

K.O.