Software Tuning, Performance Optimization & Platform Monitoring
Discussion regarding monitoring and software tuning methodologies, Performance Monitoring Unit (PMU) of Intel microprocessors, and platform updating.

Some inconsistency in QPI traffic monitoring for TxL_FLITS_G0

Alexander_Alexeev
257 Views

Hello 

I did couple experements measuring QPI trafic with customized version of PCM. 

I measured DATA, NONDATA and IDLE traffic over QPI link using two events RxL_FLITS_G0 and TxL_FLITS_G0. My expectations were that results of those events should be semetric. That is amount of data/nondata flits sent should be equal to amount of flits received and Idle rate should be similar for both reciver and transmitter. And also sum of all three would give maximum available bandwidth from QPI about 14 GB/sec. 

Data and nodata results are consistent but for idle in case of collecting with TxL_FLITS_G0 consistency is lost.

For each link on idle Windows host I have the following numbers 

DATA: 70 MB/sec

NONDATA: 46 MB/sec

IDLE: 88 MB/sec - while here I expected to see about 14000 MB/sec

I Use the following condifuration to collect IDLE flits count 

enable_counter(0);
configure_counter(0, Q_P_PCI_PMON_CTL_EN + Q_P_PCI_PMON_CTL_EVENT(0x00) + Q_P_PCI_PMON_CTL_UMASK(1));

Any ideas?

Thanks,

Alexander

 

0 Kudos
9 Replies
Patrick_F_Intel1
Employee
257 Views

Hello Alexander,

I'm guessing that the QPI system has some power savings mode built in to avoid sending pointless empty messages... but that is a guess at this point. It is also possible that the events aren't configured correctly or something else (the list is pretty big).

What sort of system (cpuid brand string, number of sockets, etc) are you using? Version of PCM?

Pat

0 Kudos
Alexander_Alexeev
257 Views

Patric, the system is dual socket E5-2680 and I use my own CLI tool derived from PCM. So version number is irrelevant :) 

The power saving is doubtful as for RxL monitoring result is consistent   

0 Kudos
Roman_D_Intel
Employee
257 Views

And also sum of all three would give maximum available bandwidth from QPI about 14 GB/sec. 

this is wrong because the link can be completely shutdown (in L1 power-saving QPI state) not sending any flits at all (no idle filts either).

0 Kudos
Roman_D_Intel
Employee
257 Views
The link power states are per link and per direction. So for example the tx direction could send idle flits but the rx direction can be shutdown for some time not having any flits.
0 Kudos
Roman_D_Intel
Employee
257 Views
The link power states are per link and per direction. So for example the tx direction could send idle flits but the rx direction can be shutdown for some time not having any flits.
0 Kudos
Roman_D_Intel
Employee
257 Views
The link power states are per link and per direction. So for example the tx direction could send idle flits but the rx direction can be shutdown for some time not having any flits.
0 Kudos
Roman_D_Intel
Employee
257 Views
The link power states are per link and per direction. So for example the tx direction could send idle flits but the rx direction can be shutdown for some time not having any flits.
0 Kudos
Roman_D_Intel
Employee
257 Views
The link power states are per link and per direction. So for example the tx direction could send idle flits but the rx direction can be shutdown for some time not having any flits.
0 Kudos
Roman_D_Intel
Employee
257 Views
The link power states are per link and per direction. So for example the tx direction could send idle flits but the rx direction can be shutdown for some time not having any flits.
0 Kudos
Reply