|
On Sun, 23 Jan 2000, Julian Anastasov wrote:
>
> Hi Joe,
>
> On Sun, 23 Jan 2000, Joseph Mack wrote:
>
> > > What means throughput 120 Mbps for the real servers in the table?
> > > Is the limit 100Mbps for the NICs? Input+Output in Full-Duplex ? I don't
> > > understand something in this table. You report that the throughput from
> > > the client to the real server (directly?) is 50Mbps. If the size of the
> > > request = size of the answer, the max throughput can be 50Mbps (reported
> > > in the real server, half-duplex).
> >
> > Yes this is a puzzle and I don't understand it myself. The connection
> > client-director is 70Mbps by netpipe for a single netpipe run. I assume
> > that means 70Mbps in each direction and that the total bit rate is
> > 140Mbps. I assume that the cards are all limited to 100Mbps by all having
> > to use the same clock speed so that they can talk to each other. I can
> > imagine that some cards might have slower average throughput because of
> > buffer filling etc, but it's hard to imagine a situation where the speed
> > would go above 100Mbps. I don't understand then why the client then can
> > register a total of 120Mbps on netpipe (6 windows each registering
> > 20Mbps). I should do the same test connecting the client directly to the
> > director running 6 netpipe sessions.
Here is yesterday's table with an extra column added, the direct
connection from client to director (c-d). It would seem that you can't get
get more than 100Mbps through a single cable (the connection from client
to director) with netpipe. The director was heavily loaded in this test
(load average >4) for 6 connections, since it has to generate the replies.
By comparison when configured for VS-DR, the director has little load.
I will redo the tests that give results >100Mbps. I am beginning to think
that netpipe may not be a good test - each window is sending its packet
and measuring the throughput for its packet. If because of contention for
resources, a netpipe process has to wait before starting to send a packet
and this wait is not included in the timing, then the measured throughput
will be higher than the real throughput. I will write to the netpipe
person and ask about this.
In the meantime I will try the tests with ftp and measure the total
time.
Joe
Results:
Connection direct direct VS-NAT VS-DR
(c-r) (c-d) (c-lvs) (c-lvs)
Throughput(Mbps)
targets
1 50 70 60 50
2 80 90 68 80
3 95 100 66 100
4 100 100 68 100
5 125 105 75 125
6 120 100 72 120
--
Joseph Mack mack@xxxxxxxxxxx
----------------------------------------------------------------------
LinuxVirtualServer.org mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
To unsubscribe, e-mail: lvs-users-unsubscribe@xxxxxxxxxxxxxxxxxxxxxx
For additional commands, e-mail: lvs-users-help@xxxxxxxxxxxxxxxxxxxxxx
|
