[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Xen-users] Large server, Xen limitations


  • To: xen-users@xxxxxxxxxxxxxxxxxxx
  • From: Steve Thompson <smt@xxxxxxxxxxxx>
  • Date: Fri, 2 Jan 2009 12:51:18 -0500 (EST)
  • Delivery-date: Fri, 02 Jan 2009 09:51:58 -0800
  • List-id: Xen user discussion <xen-users.lists.xensource.com>

Some quick timings, fwiw.

For those that missed the season premiere, I am compiling an application consisting of about 500K lines of C and C++ code in 1200 source files with about 100 makefiles, using GNU make. Dom0 is a Dell PE2900 with two quad-core 2.33 GHz processors and 24 GB memory; all file systems are RAID-1 SATA 7.2K rpm; xen is 3.0.3. Sources are read from an NFS mount in each case, where Dom0 is the NFS server (except for the Dom0 test itself, which is local); objects are written to a different NFS mount. Dependencies were evaluated in an earlier step.

Timings (min:sec elapsed) are the average of two runs (from "time make"); everything was idle (except for the guest) during each run. CPU utilization in the guest, as reported by "time", was about 85% in each case. Each guest has 1 VCPU and 512 MB memory. For HVM, each qemu-dm on Dom0 consumed between 10% and 25% of one core while compiling. The ccache was cleared before each run, where applicable. There were no Linux PV guests without NFS root.

CentOS 5.2 (dom0), x86_64, 2.6.18-92.1.18.el5xen, gcc 4.1.2     4:58

Fedora 6, i686, 2.6.20, gcc 4.1.2, PV, NFS root                 5:52
Fedora 6, x86_64, 2.6.20, gcc 4.1.2, PV, NFS root               6:39

Fedora 9, i686, 2.6.25.3, gcc 4.3.0, PV, NFS root               6:21
Fedora 9, x86_64, 2.6.25.3, gcc 4.3.0, PV, NFS root             7:09

RHEL4, i686, 2.6.9-78.0.8.EL, gcc 3.4.6, HVM (file-based)       11:03
RHEL4, x86_64, 2.6.9-78.0.8.EL, gcc 3.4.6, HVM (file-based)     12:25

A re-run on Fedora 9 using the ccache'd objects reduced the time to 3:17 (64-bit) and 3:05 (32-bit). Compilation on Dom0 using -j4 gives a time of 1:28. I have not tried a guest with more than one VCPU (will do so in another episode).

Each 32-bit build is 12% faster than the corresponding 64-bit build, PV or HVM. I cannot compare these timings with the Windows timings, since the software being compiled is not the same.

Timings for running another small-memory CPU-bound application (no I/O), relative to the Dom0 performance:

        Dom0                    1.00
        Fedora 6, x86_64        0.99
        Fedora 6, i686          0.88
        Fedora 9, x86_64        1.00
        Fedora 9, i686          0.85
        RHEL4, x86_64           0.95
        RHEL4, i686             0.80

Steve

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.