|
Hi,
Thank you for sending me the benchmark result.
mrds@xxxxxxxxxxxx wrote:
> Hi,
> Thanks your LVS program and and I have some questions about LVS.
>
> Question 1:
> -----------
> It is my Linux Director benchmark result. (using the ZD's WebBench 3.0)
> The result is 2 types; one is Requests/Sec. the other is Throughput
> (Bytes/sec.)
> The Linux Virtual Server is Very Good performance !!
> By the way, How do you verify the fact of load balancing of LVS ??
> I'd like to hear your comment about it.
> I want your bechmark test results or program or URL or any clues.
>
I don't have enough equipments. However, last september I did test a load
balancer
and two real servers in a 10Mbps network, with ZDLab's WebBench 1.3. I borrowed
5 PC from my classmates, which are used as web clients. I nearly achieved 1+1=2
performance. When the total performance reachs over 800KBytes/s, the 10Mbps
network becomes the bottleneck.
I think the good testing environment should have 100Mbps switched network.
However, I don't have it now.
>
> - Note : My experimental environment (text file) and benchmark
> result(MS-Excel files) are attached.
> - Homogeneous Server :
> . Pentium2-300 ->1 Client
> . Pentium2-300 ->30 Clients
> . Round-Robin (40% increase)
> . Least-Connection (60% increase !!) Pentium2-300 * 2EA
>
Would you please explain these numbers to me? just 60% increase?
>
> Question 2:
> -----------
> Some LB companys support the adaptive load balancing method.
> I have an idea whose scenario is following.
>
> - The agent monitors load - CPU, memory, etc- in the real server
> - and calculates the appropriate load value
> - and reports the load to the Linux Director periodically
> - then Linux Director's ipvsadm adjustes the weight value of WLC or WRR.
>
> What do you think about an "agent" method ?
>
We are going to implement it.
>
> Question 3:
> -----------
> The ipvsadm program doesn't well weight adjustment in the WLC...Why?
> Mm.., in the ipvsadm source, "mc.u.vs_user.wheght" value is constant ?
>
No, it isn't constant. It is server weight, you can adjust it.
>
> -----------------------------------------------------------------------------------
> Daeho Kim
> job : full time student (BIT education center /Nerwork expert class(NX102))
> address: 391-269 ,Daebang-dong ,Dongjak-goo, Seoul, Korea(ROK)
> zipcode: 156-020
> Celluler Phone : +82-011-343-1673
> e-mail : mrds@xxxxxxxxxxxx
> -----------------------------------------------------------------------------------
>
> ------------------------------------------------------------------------
>
> client1 (203.248.108.7) client2 (203.248.108.8) controller
> (203.248.108.9)
> | (P2-300/ Win98) | (P-mmx166/ Win98) | (P-mmx200/
> WinNT)
> | | |
> +-------------------------+ | +--------------------------+
> | | |
> | | |
> =======
> H U B (5 ports)
> =======
> |
> |
> |
> --------------
> Linux Director (eth0: 203.248.108.11 / eth1:
> 192.168.1.1)
> B O X (P2-400/ Linux 2.2.9/ VS-0.4)
> --------------
> |
> |
> |
> =======
> H U B (16 ports)
> =======
> | | |
> +----------------+ | +------------------+
> | | |
> | | |
> Real Server1 | Real Server2
> (192.168.1.4) | (192.168.1.2)
> (P2-300/ Linux 2.0.36) | (SPARC4/ SunOS 5.5.1)
> (Apache) | (Apache)
> |
> |
> Real Server3
> (192.168.1.5)
> (P2-300/ Linux 2.0.36)
> (Apache)
>
> ------------------------------------------------------------------------
>
I think you should use more Win98 clients. The performance of
Win98 socket code is much lower than Linux and Solaris. A
Win98 client can never overload a Linux box for normal web
service. :)
Please try the ipvs-0.5, and test VS-Tunneling and VS-DRouting.
I will be happy to see the benchmarks.
Thanks,
Wensong
|
