|
Hi all,
After a big hacking time, new Keepalived release is now finally available
on Keepalived website. This release is focused on High performances
extensions. Lot of work has been done to optimize and reduce to the max
both scheduling jitter and CPU. The code has been profiled and tested on a
big env. The main benefit with this new release is the extended scheduling
design that now support o(1) class lookup speed using a customized hash
index. All other lookup has been extended to support same kind of o(1)
complexity. Lot of others addon has been added that push this new release
to high perf and optimized grad (I hope !). This code has been tested with
succes with a conf running over 500 VRRP instances on 500vlaned interfaces.
Due to a scheduling auto-recalibration design and with the o(1) lookup
speed, the scheduling jitter observed for this code is quite inexistant :)
Others things has been extended and optimized (cf changelog).
The ChangeLog for this release is :
2003-07-22 Alexandre Cassen <acassen@xxxxxxxxxxxx>
* keepalived-1.1.0 released.
* The release focus is : "High Performance"
* Name cleanup for the healthchecking directory. use check
instead of healthcheck to be in conformance with watchdog and
global software architecture.
* updated the SYNOPSIS file for documenting the table arg inside
virtual/static_routes declaration. You can set routes refering
to a specific TABLE-ID.
* Added a dummy debug var in the genhash declaration code to
support compilation when compilation is done with debug flag.
* Added a set flag inside the real_server declaration correctly
relfect the IPVS topology when inhibit_on_failure is used.
* fixed a daemon.h include depandency on signal.h
* VRRP : Added support to a global shared buffer for incoming
advert handling. A new buffer is no longer allocated each time
processing incoming advert, instead a shared room is used.
* VRRP : Added support to pre-allocated shared buffer for
outgoing adverts. Each vrrp instance use a 'one time'
allocated buffer instead of a 'all time' one.
* VRRP : Extended the socket pool design to support shared fd
for the outbound channel. Now, socket pool create a sending
socket and affect the fd returned to vrrp instances. This
forces instances to use a shared socket instead of creating
new socket for each outgoing adverts. The error detection
is based on the incoming socket, so that outgoing socket is
not created as long as incoming socket can not be created.
* Added support to netlink ipaddress as global keyword
"static_ipaddress".
look at doc/samples/keepalived.conf.static_ipaddress.
IP addresses specified into this block will be added during
daemon bootstrap and removed during daemon shutdown.
Differential conf parsing is enabled for this block,
removing/adding static_ipaddress can be done on the fly
sending SIGHUP signal to daemon.
* VRRP : Extended track_interface to support multiple interface
tracking. For those familiar with Nokia monitored circuit,
this extention provide the same functionality.
look at doc/samples/keepalived.conf.track_interface.
* VRRP : The VRRP instance lookup framework has been extended
to use a o(1) scheduling design. Rewrote the whole instance
lookup to use o(1) lookup instead of previous o(n^2). When
receiving incoming adverts vrrp_scheduler performs a lookup
over the VRID received to get local instance representation.
Since the internal instance representation is an non-sorted
linked list, then we run a lookup at o(n^2) complexity that
introduce lantency and scheduling jitter side effect when
runing large number of instances. To avoid this limitation
a static hash table of 255 buckets were created. Since
lookup is performed over VRID and since VRID is 8bit fixed,
then the hashkey will be VRID. In order to extend code the
hashkey is based on incoming fd too. Internally, a NIC is
represented by a 2 fds : sending socket and receiving socket.
Those fds are NIC specific so we are using them as a hash
table lookup collision resolver. With this design we can now
use the same VRID on different NICs. The collision design
is a linked list so lookup is o(n^2) but due to low number
of entries we can consider o(1) speed. But to reach best
perf, differents VRID on all instance must be used. The
design can be sumed by :
VRID hash table :
+---+---+---+---+---+---+.........+-----+
| 1 | 2 | 3 | 4 | 5 | 6 |.........| 255 |
+---+---+---+---+---+---+.........+-----+
| |
+---+ +---+
|fd3| |fd1|
+---+ +---+
|
+---+
|fd5|
+---+
This hash table is filled during configuration parsing and
VRRP instances are not duplicated but dynamically pointed
to optimize memory.
* VRRP : The VRRP synchronization group lookup has been
extended. During bootstrap a VRRP instance index is built upon
sync_group instance name. This extension speed up
synchronization since while synchronizing it perfoms the
instance index instead of lookup by instance_name. The
previous synchornization code has been rewritten to use this
'list visiting' design for FAULT/BACKUP/MASTER states
synchronization.
* VRRP : Optimized the vrrp_timer_vrid_timeout(...) to speed
up vrid lookup over timeouted fd using a one pass lookup.
* Bradley Baetz, <bradley.baetz@xxxxxxxxxxxxxxx> extended
the scheduler framework to support child process handling.
Adding support to new thread child facility for handling
child processes, and modifying the scheduling select
loop & signal handling to catch SIGCHLD, and call the
appropriate process.
* Bradley Baetz, <bradley.baetz@xxxxxxxxxxxxxxx> fixed
the misc_check healthchecker using new thread child
scheduling facility. Introduced a new keyword
"misc_timeout" to kill processes which take too long
time (default is delay_loop). SIGKILL is send to processes
if they take too long time to shutdown.
* Bradley Baetz, <bradley.baetz@xxxxxxxxxxxxxxx> extended
daemon framework to block SIGCHLD to only receive it
whn its unblocked in the scheduling loop.
* Extended healthchecker delay_loop to support long
delay (ie: >1000s).
* VRRP : Added support to a shared kernel netlink command
channel for setting ip address and routes.
* Extended the genhash code to support verbose output
selection. command arg '-v' will generate a very verbose
output.
* VRRP : Extended the logging code to select verbose log
output or not. This selection is done by passing the
'-D' option to command line while starting daemon.
By default the output is silent.
* VRRP : Extended the gratuitous ARP framework to support
shared buffer and shared socket. This increase performances
for instances owning a bunch of VIP.
* VRRP : Extended the scheduling timer computation to support
timer auto-recalibrating. While computing next instance
timer, the scheduler will substract the time taken by
previous advert handling. This provide software overhead
adaptation. The recalibration is performed over usec timer
to not pertube global scheduler.
* VRRP : Fixed a gratuitous ARP issue. Extended the
ipaddress framework to point directly to interface
reflected by netlink channel instead of storing device
index. Extended the gratuitous ARP code to use new
ipaddress structure and for sending garp over device
ipaddess belong to. Needed if you run an instance on
one device interface and set VRRP VIP on different
interface.
* Extended watchdog framework to support polling delay
selection via daemon command line. Created two new
cmdline options :
--wdog-vrrp -R Define VRRP watchdog polling
delay. (default=5s)
--wdog-check -H Define checkers watchdog
polling delay. (default=5s)
* Extended SMTP code to support bigger buffer while
processing remote mta messages.
* Erik Barker, <erikb@xxxxxxxxxxxxx> extended initscript
to support native redhat init functions.
* Extended the autoconf scripts and Makefile(s) to support
code profiling. New configure option : --enable-profile
* list library has been extended to support multi-sized list &
specific element deletion. Extended to return when list is
empty. This reduce duplicated code to test is list is empty
while processing.
* VRRP : Extended VRRP scheduler to support fd hash
table design. Speed up instance lookup while
computing instance sands. This offer o(1) design
if we consider limited number of instances per
device.
* VRRP : Extended vrrp new socket creation to replace
refreshed instance fd into fd hash table index.
* VRRP : Extended vrrp framework to support
blank virtual_ipaddress block, can be usefull
if someone want to use just the VRRP advert
as hello monitoring channel.
* Some code cleaning.
Any comments are welcome,
Best regards,
Alexandre
|
