Hello,

thank you for your advices and sorry to not have detailed infra

ismael> I'm currently benchmarking new hardware aimed to serve around
ismael> 70k users For now, our IMAP server have 13k users.

This doesn't help us help you.  Is this a new rasperry Pi 4?  Is it a
Dual CPU AMD Rzyzen with 128gb of memory and fast NVMe disks?  What is
your system setup? 

Sorry, I have two servers to bench :

- first one (a model like our current IMAP servers) is 18To HDD, 256Go RAM, 8c/16th

- second (new one aimed to serve many more customers) is 24 x 14 TO (HDD SAS), 192GB DDR4 2,6Ghz,  12c/24t - 2.4GHz/3.5GHz

OS is FreeBSD 12.2

ismael> To run imaptest, I've spwan some bench clients.

Are these tests run from remote hosts?  What kind of network are you
using?  

Yes, imaptest is running from kvm remote virtual machines in the same DC.
They are some networks hops between them, but few.

ismael> Each bench client can run imaptest with 1000 clients.
ismael> More than 1000 clients will load CPU of this bench client

ismael> imaptest command (command are chosen from usage stat on our other IMAP servers):

ismael> imaptest host=xxxxx port=xxx userfile=userfile mbox=/root/dovecot-crlf
ismael>         pass=sxxxx seed=123 clients=1000 select=194 uidfetch=94 noop=70
ismael>         status=82 append=49 fetch=276 list=12 store=19 expunge=22
ismael>         msubs=4 search=4 logout=1 delete=81 no_pipelining

ismael> With one bench client, everything runs smoothly.

ismael> # ps aux | grep dovecot | awk '{print $11,$12,$13,$14,$15,$16,$17,$18}' | sort | uniq -c
ismael>      1 anvil: [221 connections] (anvil)
ismael>    1 auth: [13 wait, 0 passdb, 0 userdb] (auth)
ismael>    1 dovecot/config
ismael>    1 dovecot/imap
ismael>   84 dovecot/imap-login
ismael>    1 dovecot/log
ismael>   20 dovecot/pop3-login
ismael>    1 grep dovecot
ismael>    1 stats: [1307 connections] (stats)

ismael> When a second instance bench instance start imaptest, clients
ismael> of first and second instance begin to stall :

ismael>  1400 stalled for 20 secs in command: 1 LOGIN "fakeuser644@mailbench...." "password"

So how is your dovecot authentication setup?  Are you using a mysql
backend?  LDAP?  Where is the server you're querying against?  Are you
running mysql on the same server you're running dovecot on?

In production, we use a remote galera cluster.
On benchmarking, for now, I use static for passdb and a file for userdb.

Are you running multiple dovecot servers with dovecot director in
front of them to help spread the load and to offer resilience if/when
a backend server fails?  

No. I'm directly benchmarking backend.

ismael> And :

ismael> # ps aux | grep dovecot | awk '{print $11,$12,$13,$14,$15,$16,$17,$18}' | sort | uniq -c
ismael>    1 anvil: [221 connections] (anvil)
ismael>    1 auth: [1227 wait, 0 passdb, 0 userdb] (auth)
ismael>    1 dovecot/config
ismael>    1 dovecot/imap
ismael>   37 dovecot/imap-login
ismael>    1 dovecot/log
ismael>   20 dovecot/pop3-login
ismael>    1 grep dovecot
ismael>    1 stats: [680 connections] (stats)

ismael> Every auth go in wait, number of connection decreases.

ismael> Using mysql or a password file give same results.

Where is mysql located?  

ismael> I have used different values for service_count with also no success.

Post your configuration details.

#doveconf -n

auth_cache_negative_ttl = 0
auth_cache_size = 100 M
auth_cache_ttl = 2 mins
auth_failure_delay = 5 secs
auth_master_user_separator = *
auth_username_chars = abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ01234567890.-_@%+
auth_username_translation = %@
auth_verbose = yes
auth_worker_max_count = 500
base_dir = /var/run/dovecot/
default_client_limit = 100000
disable_plaintext_auth = no
imap_idle_notify_interval = 30 secs
listen = xxxxxxxxxxxx
login_greeting = xxxxxxxxxxxxxxxxxx
login_trusted_networks = xxxxxxxxxxxxxxxxxxx
mail_gid = xxxx
mail_uid = xxxx
mailbox_list_index = no
namespace {
  inbox = yes
  location =
  prefix = INBOX.
  separator = .
  type = private
}
namespace {
  hidden = yes
  inbox = no
  list = no
  location =
  prefix =
  separator = .
  type = private
}
passdb {
  args = password=#hidden_use-P_to_show#
  driver = static
}
plugin {
  acl = vfile
  quota = maildir:User quota
}
protocols = imap pop3

service anvil {
  client_limit = 97000
  unix_listener anvil-auth-penalty {
    mode = 00
  }
}
service auth-worker {
  client_limit = 1
  idle_kill = 0
  process_limit = 6000000
  process_min_avail = 0
  service_count = 1
  vsz_limit = 18446744073709551615 B
}
service auth {
  client_limit = 0
  idle_kill = 0
  process_limit = 1
  process_min_avail = 1
  service_count = 0
  vsz_limit = 1000 M
}
service imap-login {
  client_limit = 26000
  process_min_avail = 16
  service_count = 0
  vsz_limit = 1 G
}
service imap {
  drop_priv_before_exec = yes
  process_limit = 10000
}
service pop3-login {
  service_count = 0
}
service pop3 {
  drop_priv_before_exec = yes
  process_limit = 10000
}
ssl = no

userdb {
  driver = passwd-file
  args = username_format=%Ln /usr/local/etc/dovecot/passwd-file
  default_fields = uid=xxxx gid=xxxx  

}
verbose_proctitle = yes
version_ignore = yes
protocol imap {
  imap_max_line_length = 64 k
  mail_max_userip_connections = 100000
  mail_plugins = quota imap_quota acl
}
protocol pop3 {
  mail_max_userip_connections = 100
  mail_plugins = quota
  pop3_logout_format = top=%t/%p, retr=%r/%b, del=%d/%m, size=%s
  pop3_uidl_format = %f
}




ismael> I think my use of imaptest could be false.

It could be.  Are you thinking that 2000 users will all be logging
into the system at the same time?  

No, except when a backend is restarted, so I put delay on imaptest command :

imaptest host=xxx port=143 userfile=userfile mbox=/root/dovecot-crlf pass=password seed=123
    clients=1000 delay=5 secs=3600 select=194 uidfetch=94 noop=70 status=82 append=49
    fetch=276 list=12 store=19 expunge=22 msubs=4 search=4 logout=10 delete=81 no_pipelining

ismael> My understanding of service auth is limited for now because
ismael> I'm quite new to Dovecot (I have previously worked with
ismael> Cyrus).

Can't really give you any hints until you tell us more about your
setup.

Thanks,
Ismaël