hfistat Command

Purpose

Displays the performance statistics of the host fabric interface.

Syntax

hfistat [-O options] [ interval [ count ] ]

hfistat [-h]

Description

The hfistat command displays performance statistics related to the host fabric interface.

The following information are the descriptions of the column headings in the output table.
Type Column name with description
HFI
HFI
Host fabric interface identifier (0, 1, ...) identifies the interface whose statistics are displayed.
packets - sent
Aggregate packet sent count (56-bit counter) is the number of packets sent to the cluster network, irrespective of the window it is sent from.
packets - imm_send
Aggregate immediate send packet sent count (56-bit counter) is the number of packets that must be sent immediately, irrespective of the window it is sent from.
packets - receive
Aggregate packet received count (56-bit counter) is the number of packets successfully received from the cluster network, irrespective of the window it is received from. The packets - receive counter increments whenever the packet received counter of any window increments.
sent_packets - fullRDMA
Aggregate full-RDMA packet sent count (56-bit counter) is the number of full-RDMA packets sent, irrespective of the window it is sent from.
sent_packets - halfRDMA
Aggregate half-RDMA packet sent count (56-bit counter) is the number of half-RDMA packets sent, irrespective of the window it is sent from. The sent_packets - halfRDMA counter does not increment for notifications.
 
sent_packets - smallRDMA
Aggregate small-RDMA packet sent count (56-bit counter) is the number of small-RDMA packets sent, irrespective of the window it is sent from. The sent_packets - smallRDMA counter does not increment for notifications.
sent_packets - ip
Aggregate IP packet sent count (56-bit counter) is the number of IP packets sent, irrespective of the window it is sent from.
sent_packets - cau
Aggregate CAU packet sent count (56-bit counter) is the number of CAU packets sent, irrespective of the window it is sent from.
sent_packets - gups
Aggregate GUPS packet sent count (56-bit counter) specifies the number of GUPS packets sent, irrespective of the window it is sent from.
dropped_packets – sending
Aggregate packet dropped from sending count (56-bit counter) specifies the number of packets dropped, and not sent by the FIFO meant for sending packets, irrespective of the window.
dropped_packets – receiving
Aggregate packet dropped from received count (56-bit counter) specifies the number of packets from the ISR that are dropped and not received, irrespective of the window it is received from.
xlat – wait
Address xlat wait count (56-bit counter) specifies the number of missed translations and that are pending. The xlat – wait register resets when you write anything to this register.
ISR
HFI
Host fabric interface identifier (0, 1, ...) identifies the interface whose statistics are displayed.
cycBlocked – sending
The cycles blocked from sending (64-bit) counter increments every 2 - 3 GHz chip cycle when the waiting flit cannot be sent over the link.
flits - sent
The flits sent (64-bit counter) is similar to the cycles blocked counter except that this counter increments each time a flit header passes over the corresponding ISR interface.
flits - dropped
The flits dropped (40-bit counter) is the number of flits that are dropped and counted whenever certain events happen such as the following items:
  1. A port’s link status bit is off.
  2. The ISR ID is not valid.
link – retries
The link level retries (24-bit counter) counter increments each time a flit is removed from the link replay buffer and sent over the link again due to an error.
NMMU
dyn_prot_cache – hits
Nest memory management unit dynamic protection cache hits.
dyn_prot_cache – misses
Nest memory management unit dynamic protection cache misses.
ATLB – hits
Nest memory management unit address translation buffer hits.
ATLB – misses
Nest memory management unit address translation buffer misses.
CAU
cycles – waiting
Cycles waiting on credit (nonindexed counter).
Window-based
HFI
Host fabric interface identifier (0, 1, ...) identifies the interface whose statistics are displayed.
Win
Window number (0, 1, 2 ...) identifies the window whose statistics are displayed.
packet_indicated – send
The packet-indicated sent count (56-bit counter) increments whenever a packet is sent that has the Packet-Indicated-Count bit in the header set.
packet_indicated – receive
The packet-indicated received count (56-bit counter) increments whenever a packet is received that has the packet-indicated-count bit in the header set, but before the packet is written to memory.
packets – sent
The packet sent count (56-bit counter) is the number of packets successfully sent to the cluster network and includes packets sent to the same HFI by sending them to the cluster network where they are wrapped back.
packets – received
The packet received count (56-bit counter) is the number of packets successfully received from the cluster network and includes packets received from the same HFI, which were wrapped back by the cluster network.
packet_dropped – sending
The packet dropped from sending count (40-bit counter) is the number of packets from a send FIFO that were dropped and not sent.
packet_dropped – receiving
The packet dropped from receiving count (40-bit counter) is the number of packets from the ISR that were dropped and not received.
immediate – send_pkts
The immediate send packet count (56-bit counter).

Flags

Item Description
Options Specifies the content and the presentation of a report. Use the Options parameter with the -O flag. 
-O option1=value1,option2=value2,option3=”value3 value4 value5"

The list of options must be comma-separated and the list of values must be enclosed in quotation marks (“ ”) and separated by spaces.

Following are the supported options with their values:
  • type = [window nonwindow hfi isr nmmu cau all ]
    The type option specifies that the hfistat command displays only the specified types of register values to be displayed. 
    Default value: hfi
    You can specify the following values to the type option:
    window
    Displays the window-based performance statistics.
    nonwindow
    Displays the nonwindow-based performance statistics.
    hfi
    Displays the performance counter values for HFI.
    isr
    Displays the integrated switch router (ISR) register values.
    nmmu
    Displays the Nest Memory Management Unit (NMMU) register values.
    cau
    Displays the Collectives Acceleration Unit (CAU) register values.
    all
    Displays all the register values.
  • display = [ raw | delta ]
    The display option is used to dump the register values. 
    Default value: none
    You can specify the following values with the display option:
    raw
    Dump the raw register values collected.
    delta
    Dump the delta values of the registers.
    Note:
    1. The hfistat tool displays formatted output when the display option is not provided.
    2. The raw and delta option values are mutually exclusive.
  • hfi = [ 0 1 … ]
    The hfi option specifies the list of host fabric interfaces for which the register values are reported. 
    Default value: All available HFIs in the system
    Note:
    1. You can specify a range of host fabric interfaces in one of the following ways: 
      hfistat –O hfi=”0 1 2 3”
      hfistat –O hfi=0-3
    2. An empty end of the list can be used to signify the last available host fabric interface. This example of the hfi option indicates the range from 1 to the last available host fabric interface. 
      hfistat –O hfi=1-
 
  • window = [ 0 1 2… ]
    The window option specifies the list of HFI window numbers for which the register values are reported. 
    Default value: All available HFI windows for the specified HFIs
    Note:
    1. You can specify a range of HFI window numbers in one of the following ways: 
      hfistat –O window=”210 211 212 213 214 215 216”
      hfistat –O window=210-216
    2. 4. An empty end of the list can be used to signify the last available HFI window number. This example specifies the window option that indicates the range from 0 to the last available HFI window. 
      hfistat –O window=0-
  • output = <filename>
    The output option specifies the output file to be used instead of stdout. 
    Default value: None
    Note: The filename is mandatory.
interval Specifies the interval in seconds for the hfistat command to collect and print statistics. If the interval parameter is not specified, the hfistat command runs with the 2-second interval.
Count Specifies the number of repetitions for the hfistat command to collect and print statistics. Use the Count parameter along with the interval option. If both the Count and the interval parameters are not specified, the hfistat command runs 10 times. If the interval parameter is specified and not the Count parameter, the hfistat command runs indefinitely.

Examples

  1. 1. To display the HFI-based performance statistics for all available HFIs by using a 2-second interval for 10 iterations, run the following command: 
    # hfistat
  2. 2. To display all the performance statistics for available HFIs including window-based performance statistics of all windows in a formatted output, run the following command: 
    # hfistat –O type=all 2 5
  3. 3. To display only the window-based performance statistics for fewer windows (0-15) in a formatted output, run the following command: 
    # hfistat –O type=window,window=0-15
  4. 4. To display the CAU register values for the HFI-1 in a formatted output for 10 samples (default) and an interval of 2 seconds (default), run the following command: 
    # hfistat –O type=cau,hfi=1
  5. 5. To display both CAU and nest memory management unit (NMMU) register values in a formatted output, run the following command: 
    # hfistat –O type=”cau nmmu”
  6. 6. To dump the raw register values of all the HFIs including window-based performance statistics for the window 0 with an interval of 2 seconds and 5 iterations, run the following command: 
    # hfistat –O display=raw,type=all,window=0 2 5
  7. 7. To dump delta values for only ISR performance counters with a 2-second interval and 5 iterations, run the following command: 
    # hfistat –O display=delta,type=isr 2 5

Files

Item Description
/usr/bin/hfistat Contains the hfistat command.