[dpdk-dev] [PATCH v2] usertools: add huge page setup script

Bruce Richardson bruce.richardson at intel.com
Fri Sep 4 17:10:14 CEST 2020


On Fri, Sep 04, 2020 at 03:58:03PM +0100, Burakov, Anatoly wrote:
> On 03-Sep-20 11:48 PM, Stephen Hemminger wrote:
> > This is an improved version of the setup of huge pages
> > bases on earlier DPDK setup. Differences are:
> >     * it autodetects NUMA vs non NUMA
> >     * it allows setting different page sizes
> >       recent kernels support multiple sizes.
> >     * it accepts a parameter in bytes (not pages).
> > 
> > If necessary the steps of clearing old settings and mounting/umounting
> > can be done individually.
> > 
> > 
> > Signed-off-by: Stephen Hemminger <stephen at networkplumber.org>
> > ---
> > v2 -- rewrite in python
> > 	The script is python3 only because supporting older versions
> > 	no longer makes any sense.
> > 
> >   usertools/hugepage-setup.py | 317 ++++++++++++++++++++++++++++++++++++
> >   1 file changed, 317 insertions(+)
> >   create mode 100644 usertools/hugepage-setup.py
> > 
> > diff --git a/usertools/hugepage-setup.py b/usertools/hugepage-setup.py
> > new file mode 100644
> > index 000000000000..8e7642428d9e
> > --- /dev/null
> > +++ b/usertools/hugepage-setup.py
> > @@ -0,0 +1,317 @@
> > +# Copyright (c) 2020 Microsoft Corporation
> > +#
> > +# Script to query and setup huge pages for DPDK applications.
> > +
> > +import sys
> > +import os
> > +import re
> > +import getopt
> > +import glob
> > +from os.path import exists, basename
> > +
> > +# convention for where to mount huge pages
> > +hugedir = '/dev/hugepages'
> 
> This isn't a "convention", this is a default systemd mountpoint.
> 
> > +
> > +# command-line flags
> > +show_flag = None
> > +reserve_kb = None
> > +clear_flag = None
> > +hugepagesize_kb = None
> > +mount_flag = None
> > +unmount_flag = None
> > +
> > +
> > +def usage():
> > +    '''Print usage information for the program'''
> > +    global hugedir
> > +    mnt = hugedir
> > +    argv0 = basename(sys.argv[0])
> > +    print("""
> > +Usage:
> > +------
> > +    %(argv0)s [options]
> > +
> > +Options:
> > +    --help, --usage:
> > +        Display usage information and quit
> > +
> > +    -s, --show:
> > +        Print the current huge page configuration.
> > +
> > +    --setup:
> > +        Simplified version of clear, umount, reserve, mount operations
> > +
> > +    -c, --clear:
> > +        Remove all huge pages
> > +
> > +    -r, --reserve:
> > +        Reserve huge pages. The size specified is in bytes, with
> > +        optional K, M or G suffix. The size must be a multiple
> > +        of the page size.
> > +
> > +    -p, --pagesize
> > +        Choose page size to use. If not specified, the default
> > +        system page size will be used.
> > +
> > +    -m, --mount
> > +        Mount the system huge page directory %(mnt)s
> > +
> > +    -u, --umount
> > +        Unmount the system huge page directory %(mnt)s
> > +
> > +
> > +Examples:
> > +---------
> > +
> > +To display current huge page settings:
> > +    %(argv0)s -s
> > +
> > +To a complete setup of with 2 Gigabyte of 1G huge pages:
> > +    %(argv0)s -p 1G --setup 2G
> > +
> > +Equivalent to:
> > +    %(argv0)s -p 1G -c -u -r 2G -m
> > +
> > +To clear existing huge page settings and umount %(mnt)s
> > +    %(argv0)s -c -u
> > +
> > +    """ % locals())
> > +
> > +
> > +def fmt_memsize(sz):
> > +    '''Format memory size in conventional format'''
> > +    sz_kb = int(sz)
> > +    if sz_kb >= 1024 * 1024:
> > +        return '{}Gb'.format(sz_kb / (1024 * 1024))
> > +    elif sz_kb >= 1024:
> > +        return '{}Mb'.format(sz_kb / 1024)
> > +    else:
> > +        return '{}Kb'.format(sz_kb)
> 
> I've lost count how many times i've had to reimplement this code, but there
> is an easier way :) Off the top of my head,
> 
> idx = log2(sz)
> # every 10th power of 2
> return '{}{}b'.format(sz, ' kMG'[int(idx) / 10])
> 
> or something close to that.
> 

Another minor nit, since these are memory sizes, not bandwidth rates, it's
bytes not bits, so the "b" should be "B" in all the prints, whatever way
it's calculated.


More information about the dev mailing list