Merge branch 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...

[mirror_ubuntu-bionic-kernel.git] / Documentation / networking / eql.txt

  EQL Driver: Serial IP Load Balancing HOWTO
  Simon "Guru Aleph-Null" Janes, simon@ncm.com
  v1.1, February 27, 1995

  This is the manual for the EQL device driver. EQL is a software device
  that lets you load-balance IP serial links (SLIP or uncompressed PPP)
  to increase your bandwidth. It will not reduce your latency (i.e. ping
  times) except in the case where you already have lots of traffic on
  your link, in which it will help them out. This driver has been tested
  with the 1.1.75 kernel, and is known to have patched cleanly with
  1.1.86.  Some testing with 1.1.92 has been done with the v1.1 patch
  which was only created to patch cleanly in the very latest kernel
  source trees. (Yes, it worked fine.)

  1.  Introduction

  Which is worse? A huge fee for a 56K leased line or two phone lines?
  It's probably the former.  If you find yourself craving more bandwidth,
  and have a ISP that is flexible, it is now possible to bind modems
  together to work as one point-to-point link to increase your
  bandwidth.  All without having to have a special black box on either
  side.


  The eql driver has only been tested with the Livingston PortMaster-2e
  terminal server. I do not know if other terminal servers support load-
  balancing, but I do know that the PortMaster does it, and does it
  almost as well as the eql driver seems to do it (-- Unfortunately, in
  my testing so far, the Livingston PortMaster 2e's load-balancing is a
  good 1 to 2 KB/s slower than the test machine working with a 28.8 Kbps
  and 14.4 Kbps connection.  However, I am not sure that it really is
  the PortMaster, or if it's Linux's TCP drivers. I'm told that Linux's
  TCP implementation is pretty fast though.--)


  I suggest to ISPs out there that it would probably be fair to charge
  a load-balancing client 75% of the cost of the second line and 50% of
  the cost of the third line etc...


  Hey, we can all dream you know...


  2.  Kernel Configuration

  Here I describe the general steps of getting a kernel up and working
  with the eql driver.	From patching, building, to installing.


  2.1.	Patching The Kernel

  If you do not have or cannot get a copy of the kernel with the eql
  driver folded into it, get your copy of the driver from
  ftp://slaughter.ncm.com/pub/Linux/LOAD_BALANCING/eql-1.1.tar.gz.
  Unpack this archive someplace obvious like /usr/local/src/.  It will
  create the following files:



       ______________________________________________________________________
       -rw-r--r-- guru/ncm	198 Jan 19 18:53 1995 eql-1.1/NO-WARRANTY
       -rw-r--r-- guru/ncm	30620 Feb 27 21:40 1995 eql-1.1/eql-1.1.patch
       -rwxr-xr-x guru/ncm	16111 Jan 12 22:29 1995 eql-1.1/eql_enslave
       -rw-r--r-- guru/ncm	2195 Jan 10 21:48 1995 eql-1.1/eql_enslave.c
       ______________________________________________________________________

  Unpack a recent kernel (something after 1.1.92) someplace convenient
  like say /usr/src/linux-1.1.92.eql. Use symbolic links to point
  /usr/src/linux to this development directory.


  Apply the patch by running the commands:


       ______________________________________________________________________
       cd /usr/src
       patch </usr/local/src/eql-1.1/eql-1.1.patch
       ______________________________________________________________________





  2.2.	Building The Kernel

  After patching the kernel, run make config and configure the kernel
  for your hardware.


  After configuration, make and install according to your habit.


  3.  Network Configuration

  So far, I have only used the eql device with the DSLIP SLIP connection
  manager by Matt Dillon (-- "The man who sold his soul to code so much
  so quickly."--) .  How you configure it for other "connection"
  managers is up to you.  Most other connection managers that I've seen
  don't do a very good job when it comes to handling more than one
  connection.


  3.1.	/etc/rc.d/rc.inet1

  In rc.inet1, ifconfig the eql device to the IP address you usually use
  for your machine, and the MTU you prefer for your SLIP lines.	One
  could argue that MTU should be roughly half the usual size for two
  modems, one-third for three, one-fourth for four, etc...  But going
  too far below 296 is probably overkill. Here is an example ifconfig
  command that sets up the eql device:



       ______________________________________________________________________
       ifconfig eql 198.67.33.239 mtu 1006
       ______________________________________________________________________





  Once the eql device is up and running, add a static default route to
  it in the routing table using the cool new route syntax that makes
  life so much easier:



       ______________________________________________________________________
       route add default eql
       ______________________________________________________________________


  3.2.	Enslaving Devices By Hand

  Enslaving devices by hand requires two utility programs: eql_enslave
  and eql_emancipate (-- eql_emancipate hasn't been written because when
  an enslaved device "dies", it is automatically taken out of the queue.
  I haven't found a good reason to write it yet... other than for
  completeness, but that isn't a good motivator is it?--)


  The syntax for enslaving a device is "eql_enslave <master-name>
  <slave-name> <estimated-bps>".  Here are some example enslavings:



       ______________________________________________________________________
       eql_enslave eql sl0 28800
       eql_enslave eql ppp0 14400
       eql_enslave eql sl1 57600
       ______________________________________________________________________





  When you want to free a device from its life of slavery, you can
  either down the device with ifconfig (eql will automatically bury the
  dead slave and remove it from its queue) or use eql_emancipate to free
  it. (-- Or just ifconfig it down, and the eql driver will take it out
  for you.--)



       ______________________________________________________________________
       eql_emancipate eql sl0
       eql_emancipate eql ppp0
       eql_emancipate eql sl1
       ______________________________________________________________________





  3.3.	DSLIP Configuration for the eql Device

  The general idea is to bring up and keep up as many SLIP connections
  as you need, automatically.


  3.3.1.  /etc/slip/runslip.conf

  Here is an example runslip.conf:















  ______________________________________________________________________
  name		sl-line-1
  enabled
  baud		38400
  mtu		576
  ducmd		-e /etc/slip/dialout/cua2-288.xp -t 9
  command	 eql_enslave eql $interface 28800
  address	 198.67.33.239
  line		/dev/cua2

  name		sl-line-2
  enabled
  baud		38400
  mtu		576
  ducmd		-e /etc/slip/dialout/cua3-288.xp -t 9
  command	 eql_enslave eql $interface 28800
  address	 198.67.33.239
  line		/dev/cua3
  ______________________________________________________________________





  3.4.	Using PPP and the eql Device

  I have not yet done any load-balancing testing for PPP devices, mainly
  because I don't have a PPP-connection manager like SLIP has with
  DSLIP. I did find a good tip from LinuxNET:Billy for PPP performance:
  make sure you have asyncmap set to something so that control
  characters are not escaped.


  I tried to fix up a PPP script/system for redialing lost PPP
  connections for use with the eql driver the weekend of Feb 25-26 '95
  (Hereafter known as the 8-hour PPP Hate Festival).  Perhaps later this
  year.


  4.  About the Slave Scheduler Algorithm

  The slave scheduler probably could be replaced with a dozen other
  things and push traffic much faster.	The formula in the current set
  up of the driver was tuned to handle slaves with wildly different
  bits-per-second "priorities".


  All testing I have done was with two 28.8 V.FC modems, one connecting
  at 28800 bps or slower, and the other connecting at 14400 bps all the
  time.


  One version of the scheduler was able to push 5.3 K/s through the
  28800 and 14400 connections, but when the priorities on the links were
  very wide apart (57600 vs. 14400) the "faster" modem received all
  traffic and the "slower" modem starved.


  5.  Testers' Reports

  Some people have experimented with the eql device with newer
  kernels (than 1.1.75).  I have since updated the driver to patch
  cleanly in newer kernels because of the removal of the old "slave-
  balancing" driver config option.


  o  icee from LinuxNET patched 1.1.86 without any rejects and was able
     to boot the kernel and enslave a couple of ISDN PPP links.

  5.1.	Randolph Bentson's Test Report






























































  From bentson@grieg.seaslug.org Wed Feb  8 19:08:09 1995
  Date: Tue, 7 Feb 95 22:57 PST
  From: Randolph Bentson <bentson@grieg.seaslug.org>
  To: guru@ncm.com
  Subject: EQL driver tests


  I have been checking out your eql driver.  (Nice work, that!)
  Although you may already done this performance testing, here
  are some data I've discovered.

  Randolph Bentson
  bentson@grieg.seaslug.org

  ---------------------------------------------------------


  A pseudo-device driver, EQL, written by Simon Janes, can be used
  to bundle multiple SLIP connections into what appears to be a
  single connection.  This allows one to improve dial-up network
  connectivity gradually, without having to buy expensive DSU/CSU
  hardware and services.

  I have done some testing of this software, with two goals in
  mind: first, to ensure it actually works as described and
  second, as a method of exercising my device driver.

  The following performance measurements were derived from a set
  of SLIP connections run between two Linux systems (1.1.84) using
  a 486DX2/66 with a Cyclom-8Ys and a 486SLC/40 with a Cyclom-16Y.
  (Ports 0,1,2,3 were used.  A later configuration will distribute
  port selection across the different Cirrus chips on the boards.)
  Once a link was established, I timed a binary ftp transfer of
  289284 bytes of data.	If there were no overhead (packet headers,
  inter-character and inter-packet delays, etc.) the transfers
  would take the following times:

      bits/sec	seconds
      345600	8.3
      234600	12.3
      172800	16.7
      153600	18.8
      76800	37.6
      57600	50.2
      38400	75.3
      28800	100.4
      19200	150.6
      9600	301.3

  A single line running at the lower speeds and with large packets
  comes to within 2% of this.  Performance is limited for the higher
  speeds (as predicted by the Cirrus databook) to an aggregate of
  about 160 kbits/sec.	The next round of testing will distribute
  the load across two or more Cirrus chips.

  The good news is that one gets nearly the full advantage of the
  second, third, and fourth line's bandwidth.  (The bad news is
  that the connection establishment seemed fragile for the higher
  speeds.  Once established, the connection seemed robust enough.)

  #lines  speed	mtu  seconds	theory  actual  %of
	 kbit/sec      duration	speed	speed	max
  3	115200  900	_	345600
  3	115200  400	18.1	345600  159825  46
  2	115200  900	_	230400
  2	115200  600	18.1	230400  159825  69
  2	115200  400	19.3	230400  149888  65
  4	57600	900	_	234600
  4	57600	600	_	234600
  4	57600	400	_	234600
  3	57600	600	20.9	172800  138413  80
  3	57600	900	21.2	172800  136455  78
  3	115200  600	21.7	345600  133311  38
  3	57600	400	22.5	172800  128571  74
  4	38400	900	25.2	153600  114795  74
  4	38400	600	26.4	153600  109577  71
  4	38400	400	27.3	153600  105965  68
  2	57600	900	29.1	115200  99410.3 86
  1	115200  900	30.7	115200  94229.3 81
  2	57600	600	30.2	115200  95789.4 83
  3	38400	900	30.3	115200  95473.3 82
  3	38400	600	31.2	115200  92719.2 80
  1	115200  600	31.3	115200  92423	80
  2	57600	400	32.3	115200  89561.6 77
  1	115200  400	32.8	115200  88196.3 76
  3	38400	400	33.5	115200  86353.4 74
  2	38400	900	43.7	76800	66197.7 86
  2	38400	600	44	76800	65746.4 85
  2	38400	400	47.2	76800	61289	79
  4	19200	900	50.8	76800	56945.7 74
  4	19200	400	53.2	76800	54376.7 70
  4	19200	600	53.7	76800	53870.4 70
  1	57600	900	54.6	57600	52982.4 91
  1	57600	600	56.2	57600	51474	89
  3	19200	900	60.5	57600	47815.5 83
  1	57600	400	60.2	57600	48053.8 83
  3	19200	600	62	57600	46658.7 81
  3	19200	400	64.7	57600	44711.6 77
  1	38400	900	79.4	38400	36433.8 94
  1	38400	600	82.4	38400	35107.3 91
  2	19200	900	84.4	38400	34275.4 89
  1	38400	400	86.8	38400	33327.6 86
  2	19200	600	87.6	38400	33023.3 85
  2	19200	400	91.2	38400	31719.7 82
  4	9600	900	94.7	38400	30547.4 79
  4	9600	400	106	38400	27290.9 71
  4	9600	600	110	38400	26298.5 68
  3	9600	900	118	28800	24515.6 85
  3	9600	600	120	28800	24107	83
  3	9600	400	131	28800	22082.7 76
  1	19200	900	155	19200	18663.5 97
  1	19200	600	161	19200	17968	93
  1	19200	400	170	19200	17016.7 88
  2	9600	600	176	19200	16436.6 85
  2	9600	900	180	19200	16071.3 83
  2	9600	400	181	19200	15982.5 83
  1	9600	900	305	9600	9484.72 98
  1	9600	600	314	9600	9212.87 95
  1	9600	400	332	9600	8713.37 90





  5.2.	Anthony Healy's Report







  Date: Mon, 13 Feb 1995 16:17:29 +1100 (EST)
  From: Antony Healey <ahealey@st.nepean.uws.edu.au>
  To: Simon Janes <guru@ncm.com>
  Subject: Re: Load Balancing

  Hi Simon,
	  I've installed your patch and it works great. I have trialed
	  it over twin SL/IP lines, just over null modems, but I was
	  able to data at over 48Kb/s [ISDN link -Simon]. I managed a
	  transfer of up to 7.5 Kbyte/s on one go, but averaged around
	  6.4 Kbyte/s, which I think is pretty cool.  :)