frustratingly slow box at 4GB, but not 1GB of memory

From: Adrian Filipi (adrian+freebsd-perf_at_ubergeeks.com)
Date: 02/03/04

  • Next message: Darcy Buskermolen: "Re: promise sx4000 card & tx4000 results over 170mb/s sustained read"
    Date: Tue, 3 Feb 2004 12:26:19 -0500 (EST)
    To: freebsd-performance@freebsd.org
    
    

    Hi folks,

            I've got a box I'm helping tune and it's not really responding as
    I'd expect.

            It's a dual 2GHz Xeon box running on a Supermicro P4DLR motherboard
    with 4GB of physical RAM. It has 4GB of swap configured on a hardware
    RAID5 configuration.

            It's running 4.9-RELEASE. It's also running Apache with PHP and
    Perl CGI's. The CGI's read/write to a MySQL DB.

            Briefly, it responds pretty quickly when I only let the kernel use
    1GB of the RAM, but gets bogged down when I let it use the full 4GB. I'm
    limiting the RAM by setting hw.physmem.

            I've tried some of the things that I could dig up in the archives,
    but the knowledge is pretty widely scattered. :-(

            I have MAXUSERS set to 96 as that gets me plenty of file
    descriptors and pid's. While tight, I believe I could get away with only
    32, as there should not be more than 300-400 processes running at any one
    time.

            I set vm.pmap.shpgperproc="300" because of warnings from the
    kernel. I'm not sure if this is the best value, but it made the warnings
    go away.

            First, upping the KVA_PAGES to 768 to achieve a 3:1 ration of
    kernel to user memory, didn't change things.

            Disabling swap seems to make a huge difference. It changes the
    time results for "time top -bu" by two orders of magnitude from 0.006s to
    0.6s! In fact, I can run the box reasonably well with 2GB of RAM with swap
    disabled.

            If I let the box boot using all of the RAM, it is very slow. e.g.
    "time top -bu >/dev/null" takes about 6 seconds.

            I am seeing large runs of time (15-30 seconds) where the system
    time usage is in the high range (75-99). Does anyone have a suggestion on
    how to determine which kernel data structures are the problems?

    thanks,

            Adrian

    --
    [ adrian@ubergeeks.com ]
    _______________________________________________
    freebsd-performance@freebsd.org mailing list
    http://lists.freebsd.org/mailman/listinfo/freebsd-performance
    To unsubscribe, send any mail to "freebsd-performance-unsubscribe@freebsd.org"
    

  • Next message: Darcy Buskermolen: "Re: promise sx4000 card & tx4000 results over 170mb/s sustained read"

    Relevant Pages

    • Re: Newbie Setting up xserver
      ... Could be, but now that I know that low memory is causing the problem, I will ... The kernel gobbles up the rest of the physical memory huh. ... Actually my RAM is split into two ... >it could be that some of the processes related to setting the timezone ...
      (comp.os.linux.setup)
    • Re: PCMCIA: unable to map memory
      ... MB RAM. ... Linux conflicts here with the pcmcia-controller. ... - A patch for the kernel (2.2. ... BadRAM by Rick van Rein (and it's successor BadMEM ...
      (comp.os.linux.portable)
    • Re: [PATCH] reserve RAM below PHYSICAL_START
      ... The "reserved RAM" can be mapped by virtualization software with to ... You're right but the relocatable kernel only works if you relocate it ... only works up to address near 1G, we can't reserve more than 1G with ...
      (Linux-Kernel)
    • Re: [BUG] 2.6.26-rc1 lost half the RAM on UltraSPARC 5
      ... Top of RAM: 0x17f46000, Total RAM: 0xff40000 ... Memory hole size: 128MB ... Kernel: Using 1 locked TLB entries for main kernel image. ... Movable zone start PFN for each node ...
      (Linux-Kernel)
    • Re: After executing MDCreateThread data abort occurs
      ... This is almost always RAM configuration, memory map, or cache problems. ... target.The code under Startup and KernelStart routines are running fine ... So the Kernel is not letting this address 0xc2000000 to be programmed. ...
      (microsoft.public.windowsce.embedded)