Collin,
I've just done a test. I switched off 'useDB4Rebuild', restarted assp and
started a rebuild. I've never tried this since I reassembled my system and
switched all the assp stuff to windows 2016 (x64).
Jan-06-18 08:15:40 RebuildSpamDB-thread rebuildspamdb-version 7.42 started
in ASSP version 2.6.2(17362)
Jan-06-18 08:15:40 warning: 'useDB4Rebuild' is NOT set to on - the rebuild
spamdb process will possibly require a very large amount of memory - but
it will run very fast!
Jan-06-18 08:15:40 Maxbytes: 20,000
Jan-06-18 08:27:45 start populating Spamdb with 975,924 records - Bayesian
check is now disabled!
Jan-06-18 08:29:06 start populating Hidden Markov Model with 1,891,536
records!
Jan-06-18 08:30:48 Total processing time: 908 second(s)
Jan-06-18 08:30:48 Total processing data: 521.15 MByte
Jan-06-18 08:30:48 Rebuild processed 18.64 files per second.
An amazing speed.
If you can accept the fact, that assp will allocate ~ 1GB more RAM - this
will be the right option for you.
Thomas
Von: "Colin Waring" <co...@dolphinict.co.uk>
An: "ASSP development mailing list" <assp-test@lists.sourceforge.net>
Datum: 05.01.2018 21:14
Betreff: Re: [Assp-test] Meltdown/Spectre
From: Thomas Eckardt [mailto:thomas.ecka...@thockar.com]
Sent: 05 January 2018 17:16
To: ASSP development mailing list <assp-test@lists.sourceforge.net>
Subject: Re: [Assp-test] Meltdown/Spectre
>>time 7,663 seconds, data 486.61 Mbyte
>This is very slow. To be honest - I'm lost for words!
>My rebuild results are:
Mine are very different
2018-01-04 22:00:00 Maxbytes: 6,000
2018-01-05 00:03:00 start populating Spamdb with 2,466,760 records -
Bayesian check is now disabled!
2018-01-05 00:07:42 Start populating Hidden Markov Model. HMM-check is
disabled for this time!
2018-01-05 00:07:43 Total processing time: 7,663 second(s)
2018-01-05 00:07:43 Total processing data: 486.61 Mbyte
2018-01-05 00:08:37 Uploading Griplist via Direct Connection
The db part looks to be fine considering the times and the extra records
that mine added. I’m wondering why I have so many more records when
Maxbytes is less and the total data is less.
My two MX have directly mounted Gluster replicas running off a Fibre
channel SAN and the rebuild only runs on one.
I have a 4GB tmpDB mounted as tmpfs:
tmpfs 4.0G 1.3G 2.8G 32%
/usr/local/assp/tmpDB
Hardware for each is Citrix XenServer 7.2 running on HP DL servers
CPU Model: Intel(R) Xeon(R) CPU E5-2640 v2 @ 2.00GHz
112GB RAM in each with 12GB allocated to each VM
Hard drives aren’t SSD but are on a 1+0 array – I forget how many drives
are in it but there’s a few. SAN is a Dell Powervault, I’d need to check
on the spec.
The VMs are Ubuntu 16.04.3 LTS
16 cores allocated in 4 socket with 4 cores per socket
Primary
top - 20:02:52 up 82 days, 3:40, 1 user, load average: 0.41, 0.18, 0.11
Tasks: 241 total, 1 running, 240 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.2 us, 0.0 sy, 0.0 ni, 99.7 id, 0.0 wa, 0.0 hi, 0.0 si,
0.0 st
KiB Mem : 12318500 total, 180648 free, 6131216 used, 6006636
buff/cache
KiB Swap: 8253436 total, 7765076 free, 488360 used. 5702644 avail Mem
Secondary/rebuild
top - 20:02:30 up 66 days, 6:59, 2 users, load average: 0.05, 0.05,
0.07
Tasks: 250 total, 1 running, 249 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.2 us, 0.1 sy, 0.0 ni, 99.7 id, 0.0 wa, 0.0 hi, 0.0 si,
0.0 st
KiB Mem : 12318500 total, 448412 free, 7276144 used, 4593944
buff/cache
KiB Swap: 8253436 total, 6071240 free, 2182196 used. 3396112 avail Mem
ASSP uses 2.3g memory
Clamd about 1G
Gluster 2.2G
Perl is v5.22.1. I believe 5.26 is coming in 18.04 LTS at the end of April
according to the release schedule. I’ll plan an upgrade sometime after
that.
The rebuild was actually quicker a while back, maybe 40m but one of the
version changes must have had an impact. I couldn’t say which though as I
only really keep an eye on the amount of data processed and the
norm/confidence.
>From my point of view the real bottleneg for the rebuild task is, that
only one core (thread) is used by this >task, even there are 12 or more
available.
>Because of this (my bad) software design, the speed of a single core
matters too much. I think about for >a while to change this. I hope, I'll
get this fixed/improved in 2018.
Improvements are always welcome to make a great product even better ?
I hope 2018 is good to you.
All the best,
Colin.
Von: "Colin Waring" <co...@dolphinict.co.uk>
An: "ASSP development mailing list" <
assp-test@lists.sourceforge.net>
Datum: 05.01.2018 16:01
Betreff: Re: [Assp-test] Meltdown/Spectre
Hi Thomas,
Thank you for the input – I do recall previously discussing ISP mode and
realising that it was for bigger deployments than ours.
We have three servers. Two handling inbound and one specifically for
Office 365 relaying. The two inbound probably do about 50,000 messages per
day between them according to infostats.
CPU Usage on both frontends is 1.62% avg and 1.49% avg respectively. I
only have a single MySQL db (general load average is around 0.1 ) and I’ve
been watching the hypervisor reports on its performance. I did set up a
Gluster sync between the two frontends so they have access to the same
corpus without having to do it over the network – that helped with
performance however I’ve never been able to get the rebuild run to be
particularly quick (Last night’s was total processing time 7,663 seconds,
data 486.61 Mbyte). I haven’t brought it up here because it doesn’t really
have much of an effect and it is likely in my setup rather than an ASSP
issue.
So I think I’ll get away with it on my setup, hopefully this information
will be helpful to other people who are trying to figure out if they’ll be
impacted.
All the best,
Colin Waring.
From: Thomas Eckardt [mailto:thomas.ecka...@thockar.com]
Sent: 05 January 2018 13:49
To: ASSP development mailing list <assp-test@lists.sourceforge.net>
Subject: Re: [Assp-test] Meltdown/Spectre
I remember an ISP issue, who used 10 assp instances with one enterprise
MySQL backend cluster, sharing all tables for all instances.
In havy workload times (100.000 or even more mails per hour), the MySQL
server was brought to its end - no matter how many physical resouces were
made available. Even holding the complete assp DB in the DB-server RAM has
not solved the problem.
With 100.000 mails per hour and ~50 DB queries per mail (HMMdb and
spamDB), the DB server has to process at least 5 million queries in one
hour.
If we exclude HMMdb and spamDB, depending on the configuration, there can
be additionaly 10 to 20 DB queries per mail (for all the other DB-tables).
Even this can lead in to a very high DB workload!
The URIBL-check can also be very resource expensive (read and write !!!).
Assume a mail with 100 different URIs is seen the first time - 100
unsuccessfull cache DB-queries, followed by 100 DNS queries, followed by
100 cache DB-writes.
To prevent this issue, assp V2 has a buildin ISP mode for HMMdb and
spamDB.
In short:
- the corpus of all instances is synchronized to a master instance (rsync
for example)
- HMMdb and spamDB are hold in memory in each instance and each worker
- HMMdb and spamDB are build on the master system and are distributed as
files to all other instances using an external script (methode of your
choice)
- all other tables are shared traditionaly - but each instance uses a
configurable DB cache to prevent repeated DB-queries for the same results
(for example IP checks, helo ....)
This ISP mode requires at least 16GB RAM per instance, if a maximum of 15
SMTP workers is used. Using more than 15 workers in an instance, produces
a large overhead without any performance improvement.
Collin, I don't know the workload and configuration of your systems - but
the math is simple.
An possible solution between the standard mode and the ISP mode can be:
- each assp instance has its own DB backend
- all DB-backends are bidirectional synchronized (asynchron) to a
DB-master-server-cluster
Depending on the overall workload, the DB-master-server-cluster must be an
enterprise cluster or something like that.
If we assume 10 assp instances, each record change in one instance will
lead in to one store and nine write sync ops at the master cluster!
If we assume five DB-write ops per mail -> 100 000 mail/h in all instances
-> 500 000 store ops/h + 4.5M sync ops/h at the master cluster.
Yes - the workload at the cluster will be very high, but it is no longer
time critical and will balance over all the time.
The disadvantage is, that the tables in all instances are never 100%
sychron and the last instance "winns" in writing the same DB-record. The
async state of the tables in all DB-backends increases with the overall
workload.
You may also think about a ring synchronization between the 10 assp
DB-backends. The cluster will not be required and the DB-backends will
have a manageable workload - but the delay of syncing a single record and
the data inconsitency over all instances will be increased.
Thomas
Von: "Colin Waring" <co...@dolphinict.co.uk>
An: "ASSP development mailing list" <
assp-test@lists.sourceforge.net>
Datum: 05.01.2018 10:45
Betreff: [Assp-test] Meltdown/Spectre
Hi All,
I’m wondering if anyone has updated their ASSP/db backends and monitored
the performance impact yet.
I’m currently working on assessing just how bad this is going to be with
how many systems I’ve got to coordinate hypervisor/OS/microcode updates on
so I’m checking around with everyone to see who’s already got some
answers.
All the best,
Colin Waring.
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Assp-test mailing list
Assp-test@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/assp-test
DISCLAIMER:
*******************************************************
This email and any files transmitted with it may be confidential, legally
privileged and protected in law and are intended solely for the use of the
individual to whom it is addressed.
This email was multiple times scanned for viruses. There should be no
known virus in this email!
*******************************************************
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Assp-test mailing list
Assp-test@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/assp-test
DISCLAIMER:
*******************************************************
This email and any files transmitted with it may be confidential, legally
privileged and protected in law and are intended solely for the use of the
individual to whom it is addressed.
This email was multiple times scanned for viruses. There should be no
known virus in this email!
*******************************************************
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Assp-test mailing list
Assp-test@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/assp-test
DISCLAIMER:
*******************************************************
This email and any files transmitted with it may be confidential, legally
privileged and protected in law and are intended solely for the use of the
individual to whom it is addressed.
This email was multiple times scanned for viruses. There should be no
known virus in this email!
*******************************************************
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Assp-test mailing list
Assp-test@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/assp-test
