The current approach of distributing item checks may lead to peaks of data collection occupying all poller processes, while idling between peaks.

The following graph shows the poller busy time together with an 'Zabbix agent' item of a random number collected every second:

The peaks are caused by a lot of SNMP GET requests with an interval of five minutes. During the peaks the poller processes are too busy to process other item types.

It would be nice if Zabbix could ensure to use poller processes more efficiently. That's to say, avoiding peaks as much as possible.