Hi there, 2013/11/7 Thinker Rix <thinke...@rocketmail.com>
> Hi Michael, > > > On 2013-11-06 11:37, Michael Schuh wrote: > >> i have serval different Systems running, >> including an old 3GHz Intel Pentium D-CPU with 2GBytes ECC Memory: >> 4 Nic, throughput max (so far): 115 MBytes/s at 20k irqs (no polling >> enabled, no special tweaking) >> 1 Nic is Broadcom, 1 Nic is Intel Pro1000 Desktop Adapter, the other two >> Nic are an Intel Pro 1000 Dual Port Server Adapter. >> Memory is a bit short in this system, but it runs fine. >> > > Thank you for this interesting insight with the Pentium-D. As far as I > figure, you are having full gigabit throughput between two interfaces with > it?! That is exactly what I want to have, too and I am happy to learn that > it is possible even with "older" dual cores. > > the 115MBytes is one nic, traffic gets splitted over the other 3 nics. yes, seems it is the maximum possible. the datarate is the native (including protocol payload/overhead) datarate. so i do not expect much more. :-) > others Systems p.e. run with Core2Duo 2,66GHz (E7300) another one with a >> Pentium 2,9GHz (G2020) >> the last one i wouldn't recommend for high throughput and low latency. >> the reaction times and the latency rises up fast >> if the throughput rises or if i add some VPN-Tunnels( AES-256). >> > > Your comment about the G2020 is interesting, cince A) that is the CPU that > I was planning to go for (due to it's ECC support) and B) I can't > understand why it performs worse, than the other CPUs, especially the much > older Pentium D. > Here is the comparison: http://ark.intel.com/compare/ > 71070,36463,27518,27517 > Could that performance ditch / latency sensivity be due to it's > https://en.wikipedia.org/wiki/Smart_Cache ? I do not see any other > difference than that. > > well, i have no clue what the problem is. i can only say, what i could observe. the observations got made from the interntal network, the fastest side, directly connected. there is also no need to investigate it further. the $customer never reaches any bandwith limits. so the comparison with the older D-CPU Systems doesn't fit fully here. the E7300 should be slower than the G2020, also due to the different cache size and memory bandwith, but isn't. both are with normal memory, no ECC. Connections are fine, all nics Intel. may be, the mainboard. the System with the G2020 CPU is one out of 3, all with the same behaviour. those are $customer bought systems. so i guess, may be, the irq handling or as ever the sum of everything. the two older systems with the D-CPU are Serverboards with ECC Memory. i think this should make them even slower. Just those both systems have no VPN tunnels and under full load the D-CPU is 80% busy. if you like to compare it to them. > so i would recommend also the Corei5, the core i3 IMO comes close to a >> Pentium CPU. >> >> imself keep the Celeron CPU's far away from me. except for small embedded >> systems in the lower range. >> >> Corei7 or Xeon is a way to much for my taste and feeling. >> > > Since I can't go for the i5 with the Supermicro X9/X10 series motherboards > that I want to buy, I will either go for the Xeon - or buy the Pentium now > and upgrade to the Xeon later on, if performance should turn out to be not > enough. > well, depending on the workload, i would take an eye on the irq rate and think about polling and may be rising the kern.hz up to 2000. one can watch this by applying "systat -vmstat 1" at the console. thoughput per nic with "systat -ifstat 1". as others recommended, i would also recommend: throw via, marvell, realtek(D-Link, Via Rhine) chipsets and also via mainboard chipsets as far as you can, if it comes to high performance. i had already the craziest behaviour with those chipsets. started from stuggles with autonegotiation up to errors in the chipsets itself not depending on the drivers. > hth. >> > > Yes, thank you for your help so far! > :-) > > Regards > > Thinker Rix > _______________________________________________ > List mailing list > List@lists.pfsense.org > http://lists.pfsense.org/mailman/listinfo/list >
_______________________________________________ List mailing list List@lists.pfsense.org http://lists.pfsense.org/mailman/listinfo/list
