mirror of
http://git.haproxy.org/git/haproxy.git/
synced 2024-12-11 14:05:12 +00:00
d008930265
Released version 2.2-dev6 with the following main changes : - BUG/MINOR: ssl: memory leak when find_chain is NULL - CLEANUP: ssl: rename ssl_get_issuer_chain to ssl_get0_issuer_chain - MINOR: ssl: rework add cert chain to CTX to be libssl independent - BUG/MINOR: peers: init bind_proc to 1 if it wasn't initialized - BUG/MINOR: peers: avoid an infinite loop with peers_fe is NULL - BUG/MINOR: peers: Use after free of "peers" section. - CI: github actions: add weekly h2spec test - BUG/MEDIUM: mux_h1: Process a new request if we already received it. - MINOR: build: Fix build in mux_h1 - CLEANUP: remove obsolete comments - BUG/MEDIUM: dns: improper parsing of aditional records - MINOR: ssl: skip self issued CA in cert chain for ssl_ctx - MINOR: listener: add so_name sample fetch - MEDIUM: stream: support use-server rules with dynamic names - MINOR: servers: Add a counter for the number of currently used connections. - MEDIUM: connections: Revamp the way idle connections are killed - MINOR: cli: add a general purpose pointer in the CLI struct - MINOR: ssl: add a list of bind_conf in struct crtlist - REORG: ssl: move SETCERT enum to ssl_sock.h - BUG/MINOR: ssl: ckch_inst wrongly inserted in crtlist_entry - REORG: ssl: move some functions above crtlist_load_cert_dir() - MINOR: ssl: use crtlist_free() upon error in directory loading - MINOR: ssl: add a list of crtlist_entry in ckch_store - MINOR: ssl: store a ptr to crtlist in crtlist_entry - MINOR: ssl/cli: update pointer to store in 'commit ssl cert' - MEDIUM: ssl/cli: 'add ssl crt-list' command - REGTEST: ssl/cli: test the 'add ssl crt-list' command - BUG/MINOR: ssl: entry->ckch_inst not initialized - REGTEST: ssl/cli: change test type to devel - REGTEST: make the PROXY TLV validation depend on version 2.2 - CLEANUP: assorted typo fixes in the code and comments - BUG/MINOR: stats: Fix color of draining servers on stats page - DOC: internals: Fix spelling errors in filters.txt - MINOR: connections: Don't mark conn flags 0x00000001 and 0x00000002 as unused. - REGTEST: make the unique-id test depend on version 2.0 - BUG/MEDIUM: dns: Consider the fact that dns answers are case-insensitive - MINOR: ssl: split the line parsing of the crt-list - MINOR: ssl/cli: support filters and options in add ssl crt-list - MINOR: ssl: add a comment above the ssl_bind_conf keywords - REGTEST: ssl/cli: tests options and filters w/ add ssl crt-list - REGTEST: ssl: pollute the crt-list file - BUG/CRITICAL: hpack: never index a header into the headroom after wrapping - BUG/MINOR: protocol_buffer: Wrong maximum shifting. - CLEANUP: src/fd.c: mask setsockopt with DISGUISE - BUG/MINOR: ssl/cli: initialize fcount int crtlist_entry - REGTEST: ssl/cli: add other cases of 'add ssl crt-list' - CLEANUP: assorted typo fixes in the code and comments - DOC: management: add the new crt-list CLI commands - BUG/MINOR: ssl/cli: fix spaces in 'show ssl crt-list' - MINOR: ssl/cli: 'del ssl crt-list' delete an entry - MINOR: ssl/cli: replace dump/show ssl crt-list by '-n' option - CI: use better SSL library definition - CI: travis-ci: enable DEBUG_STRICT=1 for CI builds - CI: travis-ci: upgrade openssl to 1.1.1f - MINOR: ssl: improve the errors when a crt can't be open - CI: cirrus-ci: rename openssl package after it is renamed in FreeBSD - CI: adopt openssl download script to download all versions - BUG/MINOR: ssl/cli: lock the ckch structures during crt-list delete - MINOR: ssl/cli: improve error for bundle in add/del ssl crt-list - MINOR: ssl/cli: 'del ssl cert' deletes a certificate - BUG/MINOR: ssl: trailing slashes in directory names wrongly cached - BUG/MINOR: ssl/cli: memory leak in 'set ssl cert' - CLEANUP: ssl: use the refcount for the SSL_CTX' - CLEANUP: ssl/cli: use the list of filters in the crtlist_entry - BUG/MINOR: ssl: memleak of the struct cert_key_and_chain - CLEANUP: ssl: remove a commentary in struct ckch_inst - MINOR: ssl: initialize all list in ckch_inst_new() - MINOR: ssl: free instances and SNIs with ckch_inst_free() - MINOR: ssl: replace ckchs_free() by ckch_store_free() - BUG/MEDIUM: ssl/cli: trying to access to free'd memory - MINOR: ssl: ckch_store_new() alloc and init a ckch_store - MINOR: ssl: crtlist_new() alloc and initialize a struct crtlist - REORG: ssl: move some free/new functions - MINOR: ssl: crtlist_entry_{new, free} - BUG/MINOR: ssl: ssl_conf always set to NULL on crt-list parsing - MINOR: ssl: don't alloc ssl_conf if no option found - BUG/MINOR: connection: always send address-less LOCAL PROXY connections - BUG/MINOR: peers: Incomplete peers sections should be validated. - MINOR: init: report in "haproxy -c" whether there were warnings or not - MINOR: init: add -dW and "zero-warning" to reject configs with warnings - MINOR: init: report the compiler version in haproxy -vv - CLEANUP: assorted typo fixes in the code and comments - MINOR: init: report the haproxy version and executable path once on errors - DOC: Make how "option redispatch" works more explicit - BUILD: Makefile: add linux-musl to TARGET - CLEANUP: assorted typo fixes in the code and comments - CLEANUP: http: Fixed small typo in parse_http_return - DOC: hashing: update link to hashing functions
84 lines
4.4 KiB
Plaintext
84 lines
4.4 KiB
Plaintext
2013/11/20 - How hashing works internally in haproxy - maddalab@gmail.com
|
|
|
|
This document describes how Haproxy implements hashing both map-based and
|
|
consistent hashing, both prior to versions 1.5 and the motivation and tests
|
|
that were done when providing additional options starting in version 2.2
|
|
|
|
A note on hashing in general, hash functions strive to have little
|
|
correlation between input and output. The heart of a hash function is its
|
|
mixing step. The behavior of the mixing step largely determines whether the
|
|
hash function is collision-resistant. Hash functions that are collision
|
|
resistant are more likely to have an even distribution of load.
|
|
|
|
The purpose of the mixing function is to spread the effect of each message
|
|
bit throughout all the bits of the internal state. Ideally every bit in the
|
|
hash state is affected by every bit in the message. And we want to do that
|
|
as quickly as possible simply for the sake of program performance. A
|
|
function is said to satisfy the strict avalanche criterion if, whenever a
|
|
single input bit is complemented (toggled between 0 and 1), each of the
|
|
output bits should change with a probability of one half for an arbitrary
|
|
selection of the remaining input bits.
|
|
|
|
To guard against a combination of hash function and input that results in
|
|
high rate of collisions, haproxy implements an avalanche algorithm on the
|
|
result of the hashing function. In all versions 1.4 and prior avalanche is
|
|
always applied when using the consistent hashing directive. It is intended
|
|
to provide quite a good distribution for little input variations. The result
|
|
is quite suited to fit over a 32-bit space with enough variations so that
|
|
a randomly picked number falls equally before any server position, which is
|
|
ideal for consistently hashed backends, a common use case for caches.
|
|
|
|
In all versions 1.4 and prior Haproxy implements the SDBM hashing function.
|
|
However tests show that alternatives to SDBM have a better cache
|
|
distribution on different hashing criteria. Additional tests involving
|
|
alternatives for hash input and an option to trigger avalanche, we found
|
|
different algorithms perform better on different criteria. DJB2 performs
|
|
well when hashing ascii text and is a good choice when hashing on host
|
|
header. Other alternatives perform better on numbers and are a good choice
|
|
when using source ip. The results also vary by use of the avalanche flag.
|
|
|
|
The results of the testing can be found under the tests folder. Here is
|
|
a summary of the discussion on the results on 1 input criteria and the
|
|
methodology used to generate the results.
|
|
|
|
A note of the setup when validating the results independently, one
|
|
would want to avoid backend server counts that may skew the results. As
|
|
an example with DJB2 avoid 33 servers. Please see the implementations of
|
|
the hashing function, which can be found in the links under references.
|
|
|
|
The following was the set up used
|
|
|
|
(a) hash-type consistent/map-based
|
|
(b) avalanche on/off
|
|
(c) balanche host(hdr)
|
|
(d) 3 criteria for inputs
|
|
- ~ 10K requests, including duplicates
|
|
- ~ 46K requests, unique requests from 1 MM requests were obtained
|
|
- ~ 250K requests, including duplicates
|
|
(e) 17 servers in backend, all servers were assigned the same weight
|
|
|
|
Result of the hashing were obtained across the server via monitoring log
|
|
files for haproxy. Population Standard deviation was used to evaluate the
|
|
efficacy of the hashing algorithm. Lower standard deviation, indicates
|
|
a better distribution of load across the backends.
|
|
|
|
On 10K requests, when using consistent hashing with avalanche on host
|
|
headers, DJB2 significantly out performs SDBM. Std dev on SDBM was 48.95
|
|
and DJB2 was 26.29. This relationship is inverted with avalanche disabled,
|
|
however DJB2 with avalanche enabled out performs SDBM with avalanche
|
|
disabled.
|
|
|
|
On map-based hashing SDBM out performs DJB2 irrespective of the avalanche
|
|
option. SDBM without avalanche is marginally better than with avalanche.
|
|
DJB2 performs significantly worse with avalanche enabled.
|
|
|
|
Summary: The results of the testing indicate that there isn't a hashing
|
|
algorithm that can be applied across all input criteria. It is necessary
|
|
to support alternatives to SDBM, which is generally the best option, with
|
|
algorithms that are better for different inputs. Avalanche is not always
|
|
applicable and may result in less smooth distribution.
|
|
|
|
References:
|
|
Mixing Functions/Avalanche: https://papa.bretmulvey.com/post/124027987928/hash-functions
|
|
Hash Functions: http://www.cse.yorku.ca/~oz/hash.html
|