On Wed, Aug 11, 2010 at 1:09 AM, Drunkard Zhang <gongfan193@xxxxxxxxx> wrote: > 2010/8/11 Jose Ildefonso Camargo Tolosa <ildefonso.camargo@xxxxxxxxx>: >> On Tue, Aug 10, 2010 at 10:19 PM, Drunkard Zhang <gongfan193@xxxxxxxxx> wrote: >>> 2010/8/11 Jose Ildefonso Camargo Tolosa <ildefonso.camargo@xxxxxxxxx>: >>>> Hi! >>>> >>>> On Tue, Aug 10, 2010 at 12:27 AM, Stand H <hstandit@xxxxxxxxx> wrote: >>>>> >>>>> >>>>> --- On Mon, 8/9/10, Jose Ildefonso Camargo Tolosa <ildefonso.camargo@xxxxxxxxx> wrote: >>>>> >>>>>> From: Jose Ildefonso Camargo Tolosa <ildefonso.camargo@xxxxxxxxx> >>>>>> Subject: Re: how much traffic can squid handle? >>>>>> To: "Drunkard Zhang" <gongfan193@xxxxxxxxx> >>>>>> Cc: "Stand H" <hstandit@xxxxxxxxx>, squid-users@xxxxxxxxxxxxxxx >>>>>> Date: Monday, August 9, 2010, 6:26 PM >>>>>> Hi! >>>>>> >>>>>> On Mon, Aug 9, 2010 at 8:18 PM, Drunkard Zhang <gongfan193@xxxxxxxxx> >>>>>> wrote: >>>>>> >>>>>> > >>>>>> > BTW, may bonding of multiple NICs helps on too many >>>>>> interrupts. >>>>>> > >>>>>> >>>>>> Or maybe just a good NIC, or a GOOD NIC + bonding :) >>>>>> >>>>> >>>>> Can you recommend a good NIC? >>>> >>>> Most Intel have behaved really well with me. As for Broadcom: bad >>>> luck, I had to disable most of the "hardware assistance", and thus: >>>> add more load to the server, I'm currently on a "avoid Broadcom" >>>> policy, but that could change in the future (I'll try them again >>>> sometime). >>>> >>> I got bottle on forcedeth shipped with nVidia MCP55 chipset, not got >>> problem on Intel e1000e yet, but CPU usage on Intel cores balanced >>> badly. On the other hand CPU time usage on AMD Opteron cores balanced >>> very good with same configuration, so confuse about this. >>> >> >> Yeah, e1000 have worked very well for me. >> >> Ok, so, you saw that CPU usage on Intel tends to be "inclined" to one >> of the cores? and on AMD it gets more "balanced"? >> >> Also, are talking about network related load here? or just about any >> processes running on Intel multi-core and AMD multicore. >> > With same multi-squid-instance configuration, same Linux distro, and > different hardware, AMD Opteron gets more balanced CPU usage, while on > Intel Xeon just one CPU core running out, others still too idle, about > 5%-15%. When that core runs out, simple TCP SYN check on service > failed occasional. > > I'm still trying to get this problem resolved...:-( Now I'm trying > linux-2.6.35 kernel :-). > So, the only logical conclusion is: AMD rules! :) (I actually like AMD, but I have never faced this kind of problem with Xeon, or maybe I'm just not paying attention, will make a few test myself, and see how it ends). Thanks for your answer!