#   8 Gbps recording with jive5ab [MonsterBuff: definitely 16, working towards 32]
#   verkout 1.7 - MV 14Apr2022 - Updated sections 0) 2) 3) and 4) following Min-GyuS mails
#   version 1.6 - EV 29Oct2020 - sysctl from cmd line req. ""'s, in /etc/sysctl.d/*.conf not
#   version 1.5 - EV 11Feb2019 - Eskil V.: spell out vm.vfs_cache_pressure = 50 for immediate use
#   version 1.4 - SC 19Oct2016 - Simon Casey's MonsterBuff values
#   version 1.3 - HV 14Aug2014 - fix typo in 'for cpu in $HTCORE' (c/o JonQ)
#   version 1.2 - HV 10Jun2015 - net_protocol values changed
#   version 1.1 - HV 10Jun2015 - udp_mem units are in pages not bytes
#   version 1.0 - HV initial version


# System modifications - static

# 0.) Disable hyperthreading. Either in bios or using the
#     following shell script. Note: "HTCORE" is a list of integers,
#     identifying hyper-threaded CPU cores. If left empty, no CPUs are taken
#     offline
#     14Apr2022 removed "Optional:" since for high-performance packet capture
#               the hyperthreaded cores actually *hurt* performance: the O/S
#               will see the HT cores are proper cores and schedule I/O on 
#               them, but that doesn't work at all and I/O performance
#               degrades.
HTCORE=
for cpu in $HTCORE; do
    echo 0 > /sys/devices/system/cpu/${cpu}/online;
done

# 1a.) 'sysctl' parameters for UDP/network memory
#     ------------------------------------------
#     NOTE: sysctl parameters can be made persistent/set at boot
#           by creating a file /etc/sysctl.d/flexbuff.conf
#           and inserting the requested sysctl values there
#
#     NOTE: when setting a sysctl parameter from the command line
#           that has multiple values (e.g. net.ipv4.udp_mem) 
#           require "...." around the values:
#               $> sysctl net.ipv4.udp_mem="xxx yyy zzz"
#           whilst in e.g. /etc/sysctl.d/flexbuff.conf there
#           should be no ""'s:
#               $> cat /etc/sysctl.d/flexbuff.conf
#               net.ipv4.udp_mem = xxx yyy zzz
#           Running "sysctl -p" will load the config written in the system
#           files. This will complain if you use quotemarks in the wrong place.

# Assume max rec. data rate of 8 Gbps / network interface
# Kernel HZ (time slice) = 100
#   => if thread swapped out, must buffer
#      1/100 of 8 Gbps = 1/100 of 1 Gbyte
#      = ~10 Mbyte. With a bit of overhead
#      we could allow for 32 Mbyte.
# packet size ~= 5000 [Mark5B frames over ethernet]
#   => 32e6 bytes / 5e3 bytes/packet ~= 6e3 packets
#      round to 8192 ...
#
net.core.netdev_max_backlog = 8192  
net.core.netdev_max_backlog = 1048576  # MonsterBuff @16Gbps

# maximum memory usage across all sockets; 192MB should be enough for now.
# only update the max value because jive5ab never relies on the default
# value but sets it explicitly
net.core.rmem_max = 201326592
net.core.rmem_max = 1073741824    # MonsterBuff @16Gbps

# same if we want to send at high speed: we never
# rely on the default socket write buffer size
net.core.wmem_max = 201326592

# Our recording will be done over UDP so it's important
# to tweak these settings. UDP shouldn't even *start*
# to think about memory usage until we hit the 64MB mark,
# start pressuring by 128MB and use 192MB max
# NOTE: the units of these values are system PAGES, not bytes
#       on Loonix/x86_64 pagesize seems to be 4096
#       so scale those values by that amount:
# NOTE NOTE NOTE NOTE NOTE:
#   See introductory notes above under "1a.)" about
#   when to use quotes and when not!
net.ipv4.udp_mem = 16384 32768 49152
net.ipv4.udp_mem = "536870912 805306368 1073741824"  # MonsterBuff @16Gbps

# 1b.) 'sysctl' parameters for dealing with directories
#      with *many* files/directories
#     ------------------------------------------

# By default Linux is quite happy to swap out inode/dentry
# caches to free up pages for _data_. This is nice if you
# expect to use the data just read/written.
#
# On FlexBuff where we expect lots of files/directories
# but the data flow is mostly unidirectional [either we're
# recording(writing) or reading] it becomes favourable
# to retain the inode/dentry cache because the data is
# not likely to be used soon after it appeared in memory
#
# So we'll teach the kernel to NOT free inode/dentry caches
# that easily such that "ls" and "find" &cet on the many
# mount points remain responsive
# 
# See https://www.kernel.org/doc/Documentation/sysctl/vm.txt
#     section: vfs_cache_pressure
# (Might even go lower than 50 - lower means more willingness
#  to retain the caches)
vm.vfs_cache_pressure = 50

# 1c.)  Tune minimum amount of free virtual memory
#       Do not use on MiniBuff I suppose ...
#       SC: "Aaand to always have 1GB of instantly grabbable memory..."
vm.min_free_kbytes = 1048576

# 2.) NIC interrupt settings
#     ----------------------
#
#     20Apr2022: It is important to keep *all* interrupt handling and recording
#                software on the same physical CPU.
#                IRQs and jive5ab process must ALL be scheduled to run on cores
#                of the same physical CPU.
#                If there is a choice of physical CPUs, choose the one where
#                the ethernet card(s) are connected to. If lucky you can use
#                the lstopo(1) [https://linux.die.net/man/1/lstopo] command
#                to investigate how the hardware in your system is connected.
#    
#     It is important that all ethernet interrupts are handled by the same
#     physical CPU. Also assure enough interrupts are generated.

# definitely NO irqbalancing!!!!!!!!!
# 14Apr2022: Depending on your actual O/S this may need a different
#            command - but the goal is to prevent the kernel from
#            from swapping the IRQ handler from one core to another:
#            each such task-switch to a different core inhibits irq
#            processing / invalidates caches and, in the case of
#            high IRQ rates (such as receiving MANY packets per second),
#            causes the system to occasionally miss a packet.
service irqbalance stop

#     Find all IRQs for the appropritate network card and force them to one
#     cpu. Note: this is specific to Intel ixgbe card on a dual hex-core
#     system.
for irq in `grep eth0-TxRx /proc/interrupts | awk -F: '{print $1;}'`; do
    echo 40 > /proc/irq/${irq}/smp_affinity;
done

# Set the rx-usecs parameter for interrupt generation. This sets the maximum
# interrupt rate to (1.0e6 / rx-usecs) interrupts/sec.

# Apparently lowering the interrupt rate did not make things better,
# although the CPU load was slightly lower [as was expected] in going
# from 'rx-usecs 1' to 'rx-usecs 15'. Packet receive loss was 0.001% in both
# cases. [~300 pkts lost on 30e6 packets]
#
# Packet timing at 8 Gbps @ 5000byte/packet is 200,000 pkts/sec =>
#   dt = 5us. So rx-usecs < 5 should be pointless; rx-usecs = 15 should
#   coalesce ~3 packets [thus lowering CPU load]. 
#   As said: for packet loss it didn't matter ...
ethtool -C ethX rx-usecs 1

# 14Apr2022:
# 2a.) hard-disk controller IRQ assignment
#
# The hard disk controller IRQs should also be pegged to one CPU core.
# Select a different one from the one handling the ethernet interrupts.
# In most systems the hard-disk controller driver will be mpt2sas:
for irq in `grep mpt2sas /proc/interrupts | awk -F: '{print $1;}'`; do
    echo 80 > /proc/irq/${irq}/smp_affinity;
done


# 3.) User process CPU pegging
#     -------------------------
#
#     Make sure jive5ab uses multiple cores of one (physical) CPU.
#     Experiment with left-over cores on the CPU handling the ethernet + hard-disk IRQs
#     or moving jive5ab completely to the other CPU.

# start jive5ab with appropriate cpu mask - note that "~irqmask" means you set
# it with the complement of the NIC and hard-disk irq masks [see step 2.)]; you may have to
# compute manually what the hex number is.
#
# 14Apr2022: On some systems it's called "numactl" to effect this
taskset ~irqmask jive5ab -m 3

# or for running jive5ab with process id PID
taskset -p <PID> ~irqmask


# 4.) jive5ab number of writer threads + socket buffer and file-chunk sizes
#     ---------------------------------------------------------------------
#
#     Before starting the FlexBuf recording, do the following:

# 14Apr2022: This used to say
#     "Configure the number of parallel writers to be equal that of the number of
#      FlexBuf mountpoints."
#
#     jive5ab allows to specify the number of net reader and disk writer threads to use dynamically
#     using the "record = nthread : <nReader> : <nWriter> ;" command.
#
#     Notes:
#       <nReader>  unused for network reading. A test implementation was written where
#                  multiple threads were reading from the same UDP socket [https://lwn.net/Articles/542629/]
#                  but it didn't seem to work as advertised: all network traffic was still delivered
#                  to only one of the threads. Since it complicates the code and had no observable
#                  gain this was not put in the production system.
#                  By not providing a value for <nReader> - "record = nthread : : <nWriter>" the
#                  system doesn't try to change the value.
#
#       <nWriter>  experiments have shown that on all tested systems maximum disk write performance
#                  was reached with <nWriter> = 3 or 4. Pushing the value beyond that usually
#                  degrades disk-write performance because the threads start competing for resources.
record = nthread : : <nWriter>

# Set the network socket buffer parameter (first number), the file chunk
# size (second number) and number of buffers to pre-allocate (third number).
# Note 1: 32 MB socket buffer seems optimal. <32 MB results in loss, same as
#         for > 32 MB
#         For speeds > 4Gbps probably not enough. 768M was used for lossless
#         16Gbps on MonsterBuff
# Note 2: pudp = UDP without 64-bit sequence number ("p" is for "plain"),
#          udp = UDP _with_ 64-bit sequence number prepended!
#         udpsnor = multiple udps streams (from different senders) can be
#         aggregated into one recording
# Note 3: preallocating the memory means that after "record=on" there is a
#         slight delay but it means that that memory is immediately
#         available to the recorder, thus making for a smoother recording.
#         If startup is deemed to slow, lower this value and hope for the
#         best.
# net_protocol = [p]udp : 32M : 256M : 32
net_protocol = [p]udp : 64M : 256M : 4
net_protocol = udpsnor : 768M : 128M    # MonsterBuff 16Gbps (note HUGE socket buffer, smaller chunk size)
                                        # recording 2x8Gbps FiLa10G streams in one go