On 09/25/2013 05:33 PM, Fujinaka, Todd wrote: > I am not familiar with this board. The rule of thumb that the lab support > guys told me was "fill all the blue DIMM slots" but I'm not sure if that's > completely correct. I just downloaded the manual and will let you know what > we find out.
I think I found the problem. The E5, running F19, was set to cpufreq mode of 'powersave'. It was not ramping up it's CPUs near high enough to get good performance. I changed it to 'performance' and now the cores run at 3.3+Ghz and performance is slightly better than the i7. The i7 appears to be running wide open the whole time, with it's cpufreq set to 'userspace'. We'll let things cook like this for a while, but I think the problem is resolved. Thanks, Ben > > Thanks. > > Todd Fujinaka > Software Application Engineer > Networking Division (ND) > Intel Corporation > [email protected] > (503) 712-4565 > > > -----Original Message----- > From: Ben Greear [mailto:[email protected]] > Sent: Wednesday, September 25, 2013 1:21 PM > To: Fujinaka, Todd > Cc: Waskiewicz Jr, Peter P; e1000-devel list > Subject: Re: [E1000-devel] Slower performance on E5 than i7? > > On 09/22/2013 10:38 AM, Fujinaka, Todd wrote: >> Can you tell us which motherboard you're using? This is the weekend for me >> so I'm just giving you a quick (little-research) answer, but most recent >> motherboards I've seen require 3 DIMMs per CPU or performance will be >> impacted. > > The motherboard is X9SRL-F. > > It has 4 x 4GB sticks of RAM. According to BIOS, the DIMM A-D1 is populated, > and the A-D2 are not populated. RAM is 1600 ECC. > > Should we aim to have all 8 slots populated? At least on older systems, I > recall that QPI (I think) decreased after a certain amount of RAM. If that > limitation still holds, please suggest the optimal RAM configuration that > gives us around 16GB total RAM. > > We have 2 2-port 10G ixgbe NICs, and 4 4-port igb NICs in the system, though > for much of our testing we have been using a single 2-port 10G NIC wired > back-to-back and the other NICs are just sitting there idle. > > Some of our testing was sending over 'veth' virtual network devices that will > never touch any NIC at all. > > Thanks, > Ben > >> >> Todd Fujinaka >> Software Application Engineer >> Networking Division (ND) >> Intel Corporation >> [email protected] >> (503) 712-4565 >> >> -----Original Message----- >> From: Ben Greear [mailto:[email protected]] >> Sent: Saturday, September 21, 2013 9:03 AM >> To: Waskiewicz Jr, Peter P >> Cc: e1000-devel list >> Subject: Re: [E1000-devel] Slower performance on E5 than i7? >> >> On 09/20/2013 11:14 PM, Waskiewicz Jr, Peter P wrote: >>> On Fri, 2013-09-20 at 20:51 -0700, Ben Greear wrote: >>>> We have been seeing across-the-board network performance issues with >>>> the E5 2643 v/s an I7-980. >>>> We've been trying ixgbe, igb, and 'veth' virtual devices. >>>> >>>> Driven by user space or pktgen, the E5 is usually slower by a >>>> noticeable bit. We did disable the VTD BIOS thing that really kills >>>> performance, but perhaps there are other BIOS things we should tune >>>> in the E5 system? >>>> >>>> We tried various kernels back to about 3.0 and the general slowness >>>> remains. >>>> >>>> Has anyone done similar comparisons and seen issues? >>> >>> If I'm not mistaken, the E5 has a lower max frequency than the >>> desktop parts. What I suspect is you're seeing your workloads >>> physically run faster on the I7 due to turbo pushing the core past 3 >>> GHz, but the E5 I think hovers in the 2.5-3.0 range. >> >> The E5 claims 3.3Ghz. The i7 is a bit faster, but not that much (the >> slightly slower I7-975 out performs it as well). >> >> For all our testing we are running extended loads, so I'm guessing the >> temporary turbo boost will not have much effect on our testing. >> >> >>> Other factors could be how memory is populated in both systems. Are >>> you sure all channels are populated in both systems? >> >> Both systems have 4x4GB sticks...I *think* it is fully populated. >> >> It seems to me that the E5 is more susceptible to cache issues...for >> instance if everything is *just* right, the E5 can come close to pushing >> full tx + rx on 4 x 10G ports using pktgen, but if softirqd twiddles with >> things performance will degrade significantly. The same issues are seen on >> the i7, but it seems to a much less degree. One of our engineers found some >> cache thing to disable in the E5 system's BIOS that gave another 15% >> improvement, but I don't have the details at this time. >> >>> This is just some of the quick low-hanging fruit I can think of now. >> >> I appreciate the help. Is there a current suggested replacement for the I7 >> (since they seem to be mostly end-of-life)? >> >> Thanks, >> Ben >> >>> >>> Cheers, >>> -PJ >>> >> >> >> -- >> Ben Greear <[email protected]> >> Candela Technologies Inc http://www.candelatech.com >> >> >> ---------------------------------------------------------------------- >> -------- LIMITED TIME SALE - Full Year of Microsoft Training For Just >> $49.99! >> 1,500+ hours of tutorials including VisualStudio 2012, Windows 8, >> SharePoint 2013, SQL 2012, MVC 4, more. BEST VALUE: New Multi-Library >> Power Pack includes Mobile, Cloud, Java, and UX Design. Lowest price ever! >> Ends 9/22/13. >> http://pubads.g.doubleclick.net/gampad/clk?id=64545871&iu=/4140/ostg.c >> lktrk _______________________________________________ >> E1000-devel mailing list >> [email protected] >> https://lists.sourceforge.net/lists/listinfo/e1000-devel >> To learn more about Intel® Ethernet, visit >> http://communities.intel.com/community/wired >> > > > -- > Ben Greear <[email protected]> > Candela Technologies Inc http://www.candelatech.com > -- Ben Greear <[email protected]> Candela Technologies Inc http://www.candelatech.com ------------------------------------------------------------------------------ October Webinars: Code for Performance Free Intel webinars can help you accelerate application performance. Explore tips for MPI, OpenMP, advanced profiling, and more. Get the most from the latest Intel processors and coprocessors. See abstracts and register > http://pubads.g.doubleclick.net/gampad/clk?id=60133471&iu=/4140/ostg.clktrk _______________________________________________ E1000-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel® Ethernet, visit http://communities.intel.com/community/wired
