Performance of scp with OpenSSH - 5.1p1 - openssh

This is a discussion on Performance of scp with OpenSSH - 5.1p1 - openssh ; Hello All, As the release notes of SSH-4.7 version says that a new MAC algorithm (UMAC-64 - RFC4418) was introduced with OpenSSH-4.7 which gives much better performance, I was tempted to check out the enhanced speed provided with new version. ...

+ Reply to Thread
Results 1 to 4 of 4

Thread: Performance of scp with OpenSSH - 5.1p1

  1. Performance of scp with OpenSSH - 5.1p1


    Hello All,
    As the release notes of SSH-4.7 version says that a new MAC algorithm
    (UMAC-64 - RFC4418) was introduced with OpenSSH-4.7 which gives much
    better performance, I was tempted to check out the enhanced speed
    provided with new version. So I downloaded OpenSSH-5.1p1 and build it on
    Solaris 10 with Sun Compiler CC.

    My test setup:-
    1. Two Sunfire 440 with 2 CPU (1281 MHz) and 4GB RAM.
    2. Network interface supports 1 Gbit/s (duplex)

    I tried transferring 1GB file between Node A and Node B connected using
    1GB interface using SCP. Node A can connect to Node B without password
    using public key authentication.

    With OpenSSH-5.1p1:-
    ------------------------
    Protocol Time taken rate MB/s
    Std. scp 53.7 19.3
    scp -c arcfour 40.1 26.3
    scp -c blowfish 55.5 18.6
    scp -C 123.6 16.3


    With OpenSSH-4.5p1:-
    -------------------------
    Protocol Time taken rate MB/s
    Std. scp 59.9 17.9
    scp -c arcfour 45 22.3
    scp -c blowfish 59.6 17.1
    scp -C 131 14.1

    I do not see any higher throughput (20%) with the 5.1p1 (with UMAC-64)
    than 4.5p1. The above results clearly says it is not increasing the
    performance by 20%. Did I something here?.
    Also does UMAC-64 have a impact on scp also or my understanding is
    wrong?.

    Thanks inadvance
    Senthil Prabu.S

    --
    Senthil Prabu.S


  2. RE: Performance of scp with OpenSSH - 5.1p1

    Hello David Green,

    > Although the algorithm may be 20% quicker, there are many
    > other factors involved with scp transfer speed (network
    > bandwidth/latency, harddisk speed, etc). A 20% increase in
    > algorithm performance won't necessarily translate to an
    > overall 20% increase in scp transfer speeds.


    I agree, but the test was carried on a machines where CPU idle time was
    around 90% through a dedicated 1 GB network interface directly connected
    between the two machine i.e. I tried transferring files across the
    cluster nodes using the dedicated cluster interconnects. Using the
    normal LAN, the time consumed was so much.

    Bottomline is there any way to improve the speed like making use of
    UMAC-64 etc....?

    --
    Senthil Prabu.S


  3. Re: Performance of scp with OpenSSH - 5.1p1

    On Mon, Aug 25, 2008 at 09:52:13PM +0530, S, Senthilprabu (NSN - IN/Bangalore) wrote:

    > With OpenSSH-5.1p1:-
    > ------------------------
    > Protocol Time taken rate MB/s
    > Std. scp 53.7 19.3
    > scp -c arcfour 40.1 26.3
    > scp -c blowfish 55.5 18.6
    > scp -C 123.6 16.3
    >
    >
    > With OpenSSH-4.5p1:-
    > -------------------------
    > Protocol Time taken rate MB/s
    > Std. scp 59.9 17.9
    > scp -c arcfour 45 22.3
    > scp -c blowfish 59.6 17.1
    > scp -C 131 14.1


    Let's reformat your data just a bit:

    Protocol 4.5p1 MB/s 5.1p1 MB/s increase
    ---------- ---------- ---------- ----------
    Std. scp 17.9 19.3 8%
    arcfour 22.3 26.3 18%
    blowfish 17.1 18.6 9%
    scp -C 14.1 16.3 16%

    > I do not see any higher throughput (20%) with the 5.1p1 (with UMAC-64)
    > than 4.5p1. The above results clearly says it is not increasing the
    > performance by 20%. Did I something here?.


    I'd say the 16% and 18% figures give reasonable credibility to the
    "approximately 20% faster" claim.


  4. Re: Performance of scp with OpenSSH - 5.1p1

    On Tuesday 26 August 2008, S, Senthilprabu (NSN - IN/Bangalore) wrote:
    > Hello David Green,
    >
    > > Although the algorithm may be 20% quicker, there are many
    > > other factors involved with scp transfer speed (network
    > > bandwidth/latency, harddisk speed, etc). A 20% increase in
    > > algorithm performance won't necessarily translate to an
    > > overall 20% increase in scp transfer speeds.

    >
    > I agree, but the test was carried on a machines where CPU idle time was
    > around 90% through a dedicated 1 GB network interface directly connected
    > between the two machine i.e. I tried transferring files across the
    > cluster nodes using the dedicated cluster interconnects. Using the
    > normal LAN, the time consumed was so much.
    >
    > Bottomline is there any way to improve the speed like making use of
    > UMAC-64 etc....?


    http://www.psc.edu/networking/projects/hpn-ssh/

    /Peter

    -----BEGIN PGP SIGNATURE-----
    Version: GnuPG v1.4.6 (GNU/Linux)

    iD8DBQBIvWnVWkaLDtAygc8RAhl6AJ9d4MmVJY0wfzdEGbS1eD 1l/mN+0ACfSOJi
    5/+1pg+9wiXGiWjB8b+H6eQ=
    =OZYt
    -----END PGP SIGNATURE-----


+ Reply to Thread