Like Ian mentioned the best way would be to use multithreaded poller. Since I see some people not able to compile it I think it is worth to try quick and dirty solution based on multiple copies of cmd.php.
Look at the beginning of cmd.php (v. 0.8.1) there is a line:
Code: Select all
$polling_items = db_fetch_assoc("select * from data_input_data_cache");
Here cacti gets all data sources to be polled and runs scripts or snmp queries one by one. If you can divide your data sources between let's say four different cmd.php pollers then you will get parallel polling at some extent. Number of such cmd.php pollers would depend on your machine size and number of your data sources.
Just to give you an example. Let's say we have 178 data sources. The machine is slow so you can estimate that it is able to poll about 50 data sources within 5 minutes. This means we need to have four pollers. I copy cmd.php to four copies with changed the line mentioned above:
Code: Select all
cmd1.php: $polling_items = db_fetch_assoc("select * from data_input_data_cache limit 0,50");
cmd2.php: $polling_items = db_fetch_assoc("select * from data_input_data_cache limit 50,50");
cmd3.php: $polling_items = db_fetch_assoc("select * from data_input_data_cache limit 100,50");
cmd4.php: $polling_items = db_fetch_assoc("select * from data_input_data_cache limit 150,50");
Ok... now I remove standard cmd.php from cron and put our new four pollers instead. This should work very well (however I did not try this specific solution yet
) - the benefit is parallel snmp polling (you are saving time) and parallel rrd files updates (which becomes a performance problem at some point too). The only thing I am not sure is logging - cacti may have problem writing to the same log in parallel.
Now about snmpget. SNMP standards allow for several OIDs in one snmpget query. This however is not implemented neither in snmpget tool we are using nor in php internal snmp support. The only way to speed up snmp polling in some cases is to use snmpbulkwalk. I for example noticed that I query almost all interfaces of given device with snmpget one by one. It takes time (especially for WAN links). If I query all interface table with snmpbulkwalk at once it takes sometimes 40 times less (!). What can be done is to separate snmp polling from cacti. Such separated poller could poll devices via snmp few snmpbulkwalk queries and put the results into some mysql table. Cacti from the other side could just read these values from that table (instead of snmp polling). Great speedup and possibility to have distributed pollers on several machines.
The last thing are hanging cmd.php pollers that Koyaanisqatsi mentioned. New cacti has the following line at the beginning of cmd.php:
Code: Select all
ini_set("max_execution_time", "0");
This means that the poller will work even a new copy is started by cron (if the first one did not finish within 5 minutes). Sometimes it's good but sometimes it results with many cmd.php copies saturating memory. You may try to change the value to 300 then after 5 minutes the poller will be killed.
- bulek
(but the box held up) and I had about 50 cmd.php's running.
at least that I could see anyway. However, since I didn't know if I was risking the integrity of my RRDs and MySQL db using cactid, I didn't let it run long enough to see any data on the graphs. It also pushed my load average up to 10 and sat there as long as cactid ran.
I'd also like to get from you a rough list of command line args and conf file options.