spamassassin (NetworkTestsLatency)
时间：2006-08-02 来源：snowtty
NetworkTestsLatency
Network Tests Latency
(original text copied from a mail to the 'users' list from Paul Shupak on 2006-01-14) One important issue often glossed over is the distinction between processor time and latency. The "net" tests in general, both DNSBL and digest (Razor, DCC and Pyzor), add significant latency to SA, but only relatively insignificant processor time. The effect of this is on a loaded server you may need more children and thus more memory, but the actual throughput will vary little with a tuned system (i.e. the proper number of children for 3.0.x or the proper choices of forking method for 3.1.x+) and the supplied default net tests. For a medium to large site it is important to consider messages per time period, not how long any one message takes to pass through the system (indeed, that is why there are multiple instances running under most setups to begin with); There is nothing wrong with allowing for 15 or 20 seconds per message to avoid Pyzor timeouts, but if make sure that the number of children will still allow the required peak message per second/minute/hour rate you need (86400 second in a day, so a 20K message/day site with 20 second DNSBL/digest timeouts will need about 5 children on the average and more during peak periods). last edited 2006-01-14 21:19:25 by JustinMason
Network Tests Latency
(original text copied from a mail to the 'users' list from Paul Shupak on 2006-01-14) One important issue often glossed over is the distinction between processor time and latency. The "net" tests in general, both DNSBL and digest (Razor, DCC and Pyzor), add significant latency to SA, but only relatively insignificant processor time. The effect of this is on a loaded server you may need more children and thus more memory, but the actual throughput will vary little with a tuned system (i.e. the proper number of children for 3.0.x or the proper choices of forking method for 3.1.x+) and the supplied default net tests. For a medium to large site it is important to consider messages per time period, not how long any one message takes to pass through the system (indeed, that is why there are multiple instances running under most setups to begin with); There is nothing wrong with allowing for 15 or 20 seconds per message to avoid Pyzor timeouts, but if make sure that the number of children will still allow the required peak message per second/minute/hour rate you need (86400 second in a day, so a 20K message/day site with 20 second DNSBL/digest timeouts will need about 5 children on the average and more during peak periods). last edited 2006-01-14 21:19:25 by JustinMason
相关阅读 更多 +
