> > > I am recording here the results from a test VME system using four VF48 waveform digitizers

Now we look at the detail of the event readout, or if you want, the real-time properties of the MIDAS 
multithreaded VME frontend program.

The benchmark system includes a TRIUMF-made VME-NIMIO32 VME trigger module which records the 
time of the trigger and provides a 20 MHz timestamp register. The frontend program is instrumented to 
save the trigger time and readout timing data into a special "trigger" bank ("VTR0"). The ROOTANA-based 
MIDAS analyzer is used to analyze this data and to make these plots.

Timing data is recorded like this:

NIM trigger signal ---> latched into the IO32 trigger time register (VTR0 "trigger time")
...
int read_event(pevent, etc) {
VTR0 "trigger time" = io32->latched_trigger_time();
VTR0 "readout start time" = io32->timestamp();
read the VF48 data
io32->release_busy();
VTR0 "readout end time" = io32->timestamp();
}

From the VTR0 time data, we compute these values:

1) "trigger latency" = "readout start time" - "trigger time" --- the time it takes us to "see" the trigger
2) "readout time" = "readout end time" - "readout start time" --- the time it takes to read the VF48 data
3) "busy time" = "readout end time" - "trigger time" --- time during which the "DAQ busy" trigger veto is 
active.
also computed is
4) "time between events" = "trigger time" - "time of previous trigger"

And plot them on the attached graphs:

1) "trigger latency" - we see average trigger latency is 5 usec with hardly any events taking more than 10 
usec (notice the log Y scale!). Also notice that there are 35 events that took longer that 100 usec (0.7% out 
of 5000 events).

So how "real time" is this? For "hard real time" the trigger latency should never exceed some maximum, 
which is determined by formal analysis or experimentally (in which case it will carry an experimental error 
bar - "response time is always less than X usec with probability 99.9...%" - the better system will have 
smaller X and more nines). Since I did not record the maximum latency, I can only claim that the 
"response time is always less than 1 sec, I am pretty sure of it".

For "soft real time" systems, such as subatomic particle physics DAQ systems, one is permitted to exceed 
that maximum response time, but "not too often". Such systems are characterized by the quantities 
derived from the present plot (mean response time, frequency of exceeding some deadlines, etc). The 
quality of a soft real time system is usually judged by non-DAQ criteria (i.e. if the DAQ for the T2K/ND280 
experiment does not respond within 20 msec, a neutrino beam spill an be lost and the experiment is 
required to report the number of lost spills to the weekly facility management meeting).

Can the trigger latency be improved by using interrupts instead of polling? Remember that on most 
hardware, the VME and PCI bus access time is around 1 usec and trigger latency of 5-10 usec corresponds 
to roughly 5-10 reads of a PCI or VME register. So there is not much room for speed up. Consider that an 
interrupt handler has to perform at least 2-3 PCI register reads (to determine the source of the interrupt 
and to clear the interrupt condition), it has to wake up the right process and do a rather slow CPU context 
switch, maybe do a cross-CPU interrupt (if VME interrupts are routed to the wrong CPU core). All this 
takes time. Then the Linux kernel interrupt latency comes into play. All this is overhead absent in pure-
polling implementations. (Yes, burning a CPU core to poll for data is wasteful, but is there any other use 
for this CPU core? With a dual-core CPU, the 1st core polls for data, the 2nd core runs mfe.c, the TCP/IP 
stack and the ethernet transmitter.)

2) "readout time" - between 7 and 8 msec, corresponding to the 50 Mbytes/sec VME block transfer rate. 
No events taking more than 10 msec. (Could claim hard real time performance here).

3) "busy time" - for the simple benchmark system it is a boring sum of plots (1) and (2). The mean busy 
time ("dead time") goes straight into the formula for computing cross-sections (if that is what you do).

4) "time between events" - provides an independent measurement of dead time - one can see that no 
event takes less than 7 msec to process and 27 events took longer than 10 msec (0.65% out of 4154 
events). If the trigger were cosmic rays instead of a pulser, this plot would also measure the cosmic ray 
event rate - one would see the exponential shape of the Poisson distribution (linear on Log scale, with the 
slope being the cosmic event rate).


K.O.

> > P.S. Observe the ever present unexplained event rate fluctuations between 130-140 event/sec.
> 
> An important aspect of optimizing your system is to keep the network traffic under control. I use GBit Ethernet between FE and BE, and make sure the switch 
> can accomodate all accumulated network traffic through its backplane. This way I do not have any TCP retransmits which kill you. Like if a single low-level 
> ethernet packet is lost due to collision, the TCP stack retransmits it. Depending on the local settings, this can be after a timeout of one (!) second, which 
> punches already a hole in your data rate. On the MSCB system actually I use UDP packets, where I schedule the retransmit myself. For a LAN, 10-100ms timeout 
> is there enough. The one second is optimized for a WAN (like between two continents) where this is fine, but it is not what you want on a LAN system. Also 
> make sure that the outgoing traffic (lazylogger) uses a different network card than the incoming traffic. I found that this also helps a lot.
> 

In typical applications at TRIUMF we do not setup a private network for the data traffic - data from VME to backend computer
and data from backend computer to DCACHE all go through the TRIUMF network.

This is justified by the required data rates - the highest data rate experiment running right now is PIENU - running
at about 10 M/s sustained, nominally April through December. (This is 20% of the data rate of the present benchmark).

The next highest data rate experiment is T2K/ND280 in Japan running at about 20 M/s (neutrino beam, data rate
is dominated by calibration events).

All other experiments at TRIUMF run at lower data rates (low intensity light ion beams), but we are planning for an experiment
that will run at 300 M/s sustained over 1 week of scheduled beam time.

But we do have the technical capability to separate data traffic from the TRIUMF network - the VME processors and
the backend computers all have dual GigE NICs.

(I did not say so, but obviously the present benchmark at 50 M/s VME to backend and 20-30 M/s from backend to HDFS is a GigE network).

(I am not monitoring the TCP loss and retransmit rates at present time)

(The network switch between VME and backend is a "the cheapest available" rackmountable 8-port GigE switch. The network between
the backend and the HDFS nodes is mostly Nortel 48-port GigE edge switches with single-GigE uplinks to the core router).

K.O.

> P.S. Observe the ever present unexplained event rate fluctuations between 130-140 event/sec.

An important aspect of optimizing your system is to keep the network traffic under control. I use GBit Ethernet between FE and BE, and make sure the switch 
can accomodate all accumulated network traffic through its backplane. This way I do not have any TCP retransmits which kill you. Like if a single low-level 
ethernet packet is lost due to collision, the TCP stack retransmits it. Depending on the local settings, this can be after a timeout of one (!) second, which 
punches already a hole in your data rate. On the MSCB system actually I use UDP packets, where I schedule the retransmit myself. For a LAN, 10-100ms timeout 
is there enough. The one second is optimized for a WAN (like between two continents) where this is fine, but it is not what you want on a LAN system. Also 
make sure that the outgoing traffic (lazylogger) uses a different network card than the incoming traffic. I found that this also helps a lot.

- Stefan

> > I am recording here the results from a test VME system using two VF48 waveform digitizers

(I now have 4 VF48 waveform digitizers, so the event rates are half of those reported before. Date rate
is up to 51 M/s - event size has doubled, per-event overhead is the same, so the effective data rate goes 
up).

This message demonstrates the effects of tuning the MIDAS system for high rate data taking.

Attached is the history plot of the event rate counters which show the real-time performance of the MIDAS 
system with better detail compared to the average event rate reported on the MIDAS status page. For an 
ideal real-time system, the event rate should be a constant, without any drop-outs.

Seen on the plot:

run 75: the periodic dropouts in the event rate correspond to the lazylogger writing data into HADOOP 
HDFS. Clearly the host computer cannot keep up with both data taking and data archiving at the same 
time. (see the output of "top" "with HDFS" and "without HDFS" below)

run 76: SYSTEM buffer size increased from 100Mbytes to 300Mbytes. Maybe there is an improvement.

run 77-78: "event_buffer_size" inside the multithreaded (EQ_MULTITHREAD) VME frontend increased from 
100Mbytes to 300Mbytes. (6 seconds of data at 50M/s). Much better, yes?

Conclusion: for improved real-time performance, there should be sufficient buffering between the VME 
frontend readout thread and the mlogger data compression thread.

For benchmark hardware, at 50M/s, 4 seconds of buffer space (100M in the SYSTEM buffer and 100M in 
the frontend) is not enough. 12 seconds of buffer space (300+300) is much better. (Or buy a faster 
backend computer).


P.S. HDFS data rate as measured by lazylogger is around 20M/s for CDH3 HADOOP and around 30M/s for 
CDH4 HADOOP.

P.S. Observe the ever present unexplained event rate fluctuations between 130-140 event/sec.


K.O.


---- "top" output during normal data taking, notice mlogger data compression consumes 99% CPU at 51 
M/s data rate.

top - 08:55:22 up 72 days, 17:00,  5 users,  load average: 2.47, 2.32, 2.27
Tasks: 206 total,   2 running, 204 sleeping,   0 stopped,   0 zombie
Cpu(s): 52.2%us,  6.1%sy,  0.0%ni, 34.4%id,  0.8%wa,  0.1%hi,  6.2%si,  0.0%st
Mem:   3925556k total,  3064928k used,   860628k free,     3788k buffers
Swap: 32766900k total,   200704k used, 32566196k free,  2061048k cached

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                                
 5826 trinat    20   0  437m 291m 287m R 97.6  7.6 636:39.63 mlogger                                                 
27617 trinat    20   0  310m 288m 288m S 24.6  7.5   6:59.28 mserver                                                 
 1806 ganglia   20   0  415m  62m 1488 S  0.9  1.6 668:43.55 gmond       


--- "top" output during lazylogger/HDFS activity. Observe high CPU use by lazylogger and fuse_dfs (the 
HADOOP HDFS client). Observe that CPU use adds up to 167% out of 200% available.

top - 08:57:16 up 72 days, 17:01,  5 users,  load average: 2.65, 2.35, 2.29
Tasks: 206 total,   2 running, 204 sleeping,   0 stopped,   0 zombie
Cpu(s): 57.6%us, 23.1%sy,  0.0%ni,  8.1%id,  0.0%wa,  0.4%hi, 10.7%si,  0.0%st
Mem:   3925556k total,  3642136k used,   283420k free,     4316k buffers
Swap: 32766900k total,   200692k used, 32566208k free,  2597752k cached

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                                
 5826 trinat    20   0  437m 291m 287m R 68.7  7.6 638:24.07 mlogger                                                 
23450 root      20   0 1849m 200m 4472 S 64.4  5.2  75:35.64 fuse_dfs                                                
27617 trinat    20   0  310m 288m 288m S 18.5  7.5   7:22.06 mserver                                                 
26723 trinat    20   0 38720  11m 1172 S 17.9  0.3  22:37.38 lazylogger                                              
 7268 trinat    20   0 1007m  35m 4004 D  1.3  0.9 187:14.52 nautilus                                                
 1097 root      20   0     0    0    0 S  0.8  0.0 101:45.55 md3_raid1   

> > > Just for completeness: Attached is the VME transfer speed I get with the SIS3100/SIS1100 interface using 
> > > 2eVME transfer. This curve can be explained exactly with an overhead of 125 us per DMA transfer and a 
> > > continuous link speed of 83 MB/sec.
>
> [with ...]  the PSI-built DRS4 board, where we implemented the 2eVME protocol in the Virtex II FPGA.

This is an interesting hardware benchmark. Do you also have benchmarks of the MIDAS system using the DRS4 (measurements
of end-to-end data rates, maximum event rate, maximum trigger rate, any tuning of the frontend program
and of the MIDAS experiment to achieve those rates, etc)?

K.O.

Hi folks:

we've been running midas-1.9.5 for a few years here at Regina.  We are now
working on a larger cosmic ray testing that requires a second ADC and second TDC
module in our Camac crate (we use the hytek1331 controller by the way).  We're
baffled as to how to set this up properly.  Specifically we have tried:

frontend.c

/* number of channels */
#define N_ADC  12 
(changed this from the old '8' to '12', and it seems to work for Lecroy 2249)

#define SLOT_ADC0   10
#define SLOT_TDC0   9
#define SLOT_ADC1   15
#define SLOT_TDC1   14

Is this the way to define the additional slots (by adding 0, 1 indices)?

Also, we were not able to get a new bank (ADC1) working, so we used a loop to
tag the second ADC values onto those of the first.

If someone has an example of how to handle multiple ADCs and TDCs and
suggestions as to where changes need to be made (header files, analyser, etc)
this would be great.

Thanks, Zisis...

P.S.  I am attaching the relevant files.

> > Just for completeness: Attached is the VME transfer speed I get with the SIS3100/SIS1100 interface using 
> > 2eVME transfer. This curve can be explained exactly with an overhead of 125 us per DMA transfer and a 
> > continuous link speed of 83 MB/sec.
> 
> What VME module is on the other end?
> 
> K.O.

The PSI-built DRS4 board, where we implemented the 2eVME protocol in the Virtex II FPGA. The same speed can be obtained with the commercial 
VME memory module CI-VME64 from Chrislin Industries (see http://www.controlled.com/vme/chinp1.html).

Stefan

> Just for completeness: Attached is the VME transfer speed I get with the SIS3100/SIS1100 interface using 
> 2eVME transfer. This curve can be explained exactly with an overhead of 125 us per DMA transfer and a 
> continuous link speed of 83 MB/sec.

What VME module is on the other end?

K.O.

> I agree. Somehow mhttpd cannot run mtransition. I am not super happy with this dependance on user $PATH settings and the inability to capture error messages 
> from attempts to start mtransition. I am now thinking in the direction of running mtransition code by forking. But remember that mlogger and the event builder also
> have to use mtransition to stop runs (otherwise they can dead-lock). So an mhttpd-only solution is not good enough...

The way to go is to make cm_transition multi-threaded. Like on thread for each client to be contacted. This way the transition can go in parallel when there are many frontend computers for example, which will speed up 
transitions significantly. In addition, cm_transition should execute a callback whenever a client succeeded or failed, so to give immediate feedback to the user. I think of something like implementing WebSockets in mhttpd for that (http://en.wikipedia.org/wiki/WebSocket).

I have this in mind since many years, but did not have time to implement it yet. Maybe on my next visit to TRIUMF?

Stefan

Just for completeness: Attached is the VME transfer speed I get with the SIS3100/SIS1100 interface using 
2eVME transfer. This curve can be explained exactly with an overhead of 125 us per DMA transfer and a 
continuous link speed of 83 MB/sec.

> I am recording here the results from a test VME system using two VF48 waveform digitizers

Note 1: data compression is about 89% (hence "data to disk" rate is much smaller than the "data from VME" rate)

Note 2: switch from VME MBLT64 block transfer to 2eVME block transfer:
- raises the VME data rate from 40 to 48 M/s
- event rate from 220/sec to 260/sec
- mlogger CPU use from 64% to about 80%

This is consistent with the measured VME block transfer rates for the VF48 module: MBLT64 is about 40 M/s, 2eVME is about 50 M/s (could be 
80 M/s if no clock cycles were lost to sync VME signals with the VF48 clocks), 2eSST is implemented but impossible - VF48 cannot drive the 
VME BERR and RETRY signals. Evil standards, grumble, grumble, grumble).

K.O.

I am recording here the results from a test VME system using two VF48 waveform digitizers and a 64-bit 
dual-core VME processor (V7865). VF48 data suppression is off, VF48 modules set to read 48 channels, 
1000 ADC samples each. mlogger data compression is enabled (gzip -1).

Event rate is about 200/sec
VME Data rate is about 40 Mbytes/sec
System is 100% busy (estimate)

System utilization of host computer (dual-core 2.2GHz, dual-channel DDR333 RAM):

(note high CPU use by mlogger for gzip compression of midas files)

top - 12:23:45 up 68 days, 20:28,  3 users,  load average: 1.39, 1.22, 1.04
Tasks: 193 total,   3 running, 190 sleeping,   0 stopped,   0 zombie
Cpu(s): 32.1%us,  6.2%sy,  0.0%ni, 54.4%id,  2.7%wa,  0.1%hi,  4.5%si,  0.0%st
Mem:   3925556k total,  3797440k used,   128116k free,     1780k buffers
Swap: 32766900k total,        8k used, 32766892k free,  2970224k cached

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                   
 5169 trinat    20   0  246m 108m  97m R 64.3  2.8  29:36.86 mlogger                                    
 5771 trinat    20   0  119m  98m  97m R 14.9  2.6 139:34.03 mserver                                    
 6083 root      20   0     0    0    0 S  2.0  0.0   0:35.85 flush-9:3                                  
 1097 root      20   0     0    0    0 S  0.9  0.0  86:06.38 md3_raid1        

System utilization of VME processor (dual-core 2.16 GHz, single-channel DDR2 RAM):

(note the more than 100% CPU use of multithreaded fevme)

top - 12:24:49 up 70 days, 19:14,  2 users,  load average: 1.19, 1.05, 1.01
Tasks: 103 total,   1 running, 101 sleeping,   1 stopped,   0 zombie
Cpu(s):  6.3%us, 45.1%sy,  0.0%ni, 47.7%id,  0.0%wa,  0.2%hi,  0.6%si,  0.0%st
Mem:   1019436k total,   866672k used,   152764k free,     3576k buffers
Swap:        0k total,        0k used,        0k free,    20976k cached

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                   
19740 trinat    20   0  177m 108m  984 S 104.5 10.9   1229:00 fevme_gef.exe                             
 1172 ganglia   20   0  416m  99m 1652 S  0.7 10.0   1101:59 gmond                                      
32353 olchansk  20   0 19240 1416 1096 R  0.2  0.1   0:00.05 top                                        
  146 root      15  -5     0    0    0 S  0.1  0.0  42:52.98 kslowd001       

Attached are the CPU and network ganglia plots from lxdaq09 (VME) and ladd02 (host).

The regular bursts of "network out" on ladd02 is lazylogger writing mid.gz files to HADOOP HDFS.

K.O.

I tried using the lazylogger "Disk" method to write into a HADOOP HDFS clustered filesystem and found a 
number of problems. I ended up replacing the lazylogger lazy_copy() function that still uses former YBOS 
code with a new lazy_disk_copy() function that uses generic fread/fwrite. Also fixed the situation where 
lazylogger cannot cleanly stop from the mhttpd "programs/stop" button while it is busy writing (the fix 
works only for the "Disk" method).

(Note that one can also use the "Script" method for writing into HDFS)

Anyhow, the new lazylogger writes into HDFS just fine and I expect that it would also work for writing into 
DCACHE using PNFS (if ever we get the SL6 PNFS working with our DCACHE servers).

Writing into our test HDFS cluster runs at about 20 MiBytes/sec for 1GB files with replication set to 3.

svn rev 5295
K.O.

> 2) cm_disconect_experiment() calls free(_net_send_buffer) but does not set its 
> value to NULL.

Set pointer to NULL after free() in these files:

M       odb.c
M       sequencer.cxx
M       mlogger.cxx
M       mhttpd.cxx
M       midas.c

svn rev 5294
K.O.

> In midas.c, the bk_delete function removes a bank by decrementing the total
> event size and then copying the remaining banks into the location of the first
> using memcpy from string.h.

Replaced some memcpy() with memmove(), including bk_delete().

svn rev 5293
K.O.

> > > Revision: r5286 
> > > Platform: Debian Linux 6.0.5 AMD64, with packages from squeeze-backports 
> 
> I found the problem only appears when I run mhttpd in scripts, whether bash or python.
> And I'm quite sure that the MIDAS environments (e.g. PATH, MIDAS_EXPTAB, MIDASSYS, etc.)
> are set in such scripts. If I start mhttpd in an xterm with or without "-D", it works
> fine.

Right. I see Debian 6.0.5 just came out hot off the presses. Would be good to fix this problem.

As a work around, can you run mhttpd without "-D", but in the background, i.e. "mhttpd -p xxx >& mhttpd.log &"?

Also what are your $PATH settings?

> So, what's the difference between invoking mhttpd directly and through a script?

As Stefan mentioned, "-D" invokes some nasty unix magic to disconnect the process from the user login session. It is 
possible that this magic breaks in the latest Debian.

MIDAS "-D" does roughly the same thing as "nohup".

K.O.

> > I found the problem only appears when I run mhttpd in scripts, whether bash or python.
> > And I'm quite sure that the MIDAS environments (e.g. PATH, MIDAS_EXPTAB, MIDASSYS, etc.)
> > are set in such scripts. If I start mhttpd in an xterm with or without "-D", it works
> > fine. So, what's the difference between invoking mhttpd directly and through a script?
> 
> When you start it with "-D", then mhttpd become a daemon. According to linux rules, it has to "cd /", so it lives in the 
> root directory, in order not to block any NFS mount/unmount. If something with the path is not correct then, mhttpd 
> cannot find mtransition then. Once I fixed that problem my moving mtransition to /usr/bin.
> 

I agree. Somehow mhttpd cannot run mtransition. I am not super happy with this dependance on user $PATH settings and the inability to capture error messages 
from attempts to start mtransition. I am now thinking in the direction of running mtransition code by forking. But remember that mlogger and the event builder also
have to use mtransition to stop runs (otherwise they can dead-lock). So an mhttpd-only solution is not good enough...

K.O.

> I found the problem only appears when I run mhttpd in scripts, whether bash or python.
> And I'm quite sure that the MIDAS environments (e.g. PATH, MIDAS_EXPTAB, MIDASSYS, etc.)
> are set in such scripts. If I start mhttpd in an xterm with or without "-D", it works
> fine. So, what's the difference between invoking mhttpd directly and through a script?

When you start it with "-D", then mhttpd become a daemon. According to linux rules, it has to "cd /", so it lives in the 
root directory, in order not to block any NFS mount/unmount. If something with the path is not correct then, mhttpd 
cannot find mtransition then. Once I fixed that problem my moving mtransition to /usr/bin.

Stefan

> > Revision: r5286 
> > Platform: Debian Linux 6.0.5 AMD64, with packages from squeeze-backports 
> > Problem:
> > After building and installation, using the script 'start_daq.sh' to start
> > 'sampleexpt'. Everything seems fine. But I cannot start a run through web. Using
> > 'odbedit' and 'mtransition' to start/stop a run works fine. So, what may cause
> > such a problem?
> 
> Well, it's mhttpd who cannot start the run, not you. So what happens when you press
> the "start run" button? Any errors in midas.log or in midas messages? Is mtransition
> in your PATH?
> 
> K.O.

I found the problem only appears when I run mhttpd in scripts, whether bash or python.
And I'm quite sure that the MIDAS environments (e.g. PATH, MIDAS_EXPTAB, MIDASSYS, etc.)
are set in such scripts. If I start mhttpd in an xterm with or without "-D", it works
fine. So, what's the difference between invoking mhttpd directly and through a script?

> Well, it's mhttpd who cannot start the run, not you. So what happens when you press
> the "start run" button? Any errors in midas.log or in midas messages? Is mtransition
> in your PATH?
After pressing "start run", there is a message displayed: "Run start requested". There
is no error in midas.log. And mtransition is actually in my PATH. I even looked into
"mhttpd.cxx" and found where "cm_transition" is called for starting a run. I have no
clue to grasp the reason.