sminfo
asked on
Increasing network trhough
We've a new server which has two network cards (1GB) incorporated. We are planning to setup NIC bonding to look for redundancy, but we want also to duplicate, is possible, the network throughput. I've the following questions:
1. How can I setup networking to duplicate or quadruplicate network through on eth0?
2. What other devices I need to achieve this?
3. Can bonding, besides solve redundancy, increase throughput on Ethernet interfaces? How? Which setup?
NOTEL I'm running Debian Linux (lenny) on it. We want also to join several other servers to this new-one, that's why we need this through
1. How can I setup networking to duplicate or quadruplicate network through on eth0?
2. What other devices I need to achieve this?
3. Can bonding, besides solve redundancy, increase throughput on Ethernet interfaces? How? Which setup?
NOTEL I'm running Debian Linux (lenny) on it. We want also to join several other servers to this new-one, that's why we need this through
ASKER CERTIFIED SOLUTION
membership
This solution is only available to members.
To access this solution, you must be a member of Experts Exchange.
The add a little, you need support for this in the switch too (802.3ad support).
You might want to add at least 4 NIC's
NIC-1 |_ Bond 1
NIC-2 |
NIC-3 |_ Bond 2
NIC-4 |
Both the bonds can can clubbed together for FAIL OVER (Similar to NIC Bonding)
and Each bond can have 2 NIC's configured for Load Balancing options.
as for bonding - specific information for given distro might be checked on their forums.
NIC-1 |_ Bond 1
NIC-2 |
NIC-3 |_ Bond 2
NIC-4 |
Both the bonds can can clubbed together for FAIL OVER (Similar to NIC Bonding)
and Each bond can have 2 NIC's configured for Load Balancing options.
as for bonding - specific information for given distro might be checked on their forums.
@Amit...
if in an agregate link (a Bond / a Channel; depending on supplier) one interface fails, the load is spread among the survivors...
So if in you example NIC-1 fails the capacity drops because only NIC-2 still works in the BOND.
The tradeof to make is do I want a GUARANTEED amount of bandwidth and redundancy you choose for failover
If you want to maximaze throughput and can work with one interface running ... (use the non-failover setup).
With multiple interfaces in a bond (not failover) the load is automaticaly spread over all interfaces, as this is comparable to striping (Raid-0) used on disks.
if in an agregate link (a Bond / a Channel; depending on supplier) one interface fails, the load is spread among the survivors...
So if in you example NIC-1 fails the capacity drops because only NIC-2 still works in the BOND.
The tradeof to make is do I want a GUARANTEED amount of bandwidth and redundancy you choose for failover
If you want to maximaze throughput and can work with one interface running ... (use the non-failover setup).
With multiple interfaces in a bond (not failover) the load is automaticaly spread over all interfaces, as this is comparable to striping (Raid-0) used on disks.
@noci
I know :)
but a simple monitoring script could compare bytes transmitted/recd (ifconfig) for each NIC over specific time frame(say 2 minutes).
He needs to figure that out :)
say if NIC 2 continues to transmit and send packets for past 2 minutes and NIC 1 lies idle - backup bond can takeover.
This will work
Out of context : multiple points of network fails where this script will work
1. Switch Port Fails
2. Network cable goes bad
3. NIC goes bad.
I know :)
but a simple monitoring script could compare bytes transmitted/recd (ifconfig) for each NIC over specific time frame(say 2 minutes).
He needs to figure that out :)
say if NIC 2 continues to transmit and send packets for past 2 minutes and NIC 1 lies idle - backup bond can takeover.
This will work
Out of context : multiple points of network fails where this script will work
1. Switch Port Fails
2. Network cable goes bad
3. NIC goes bad.
Then why not bond 4 interfaces, then you have quad throughput and triple when 1 adapter fails..?
@noci
does make sense ! any gotcha's there ?
does make sense ! any gotcha's there ?
Depending on switch config like: it needs to be one switch preferable with several blades or a stacked switch or so.
If you cannot have one switch you make it a setup like you have with spanning tree to protect youself.
If you cannot have one switch you make it a setup like you have with spanning tree to protect youself.
nice thread
I still do not know if the hardware is able to sustain more than a gigabit interface. At least, I have tested servers with 16 processors and they simply cannot sustain the load when you force users enough to fill the gigabit link.
so, my question is: is not there other bottleneck to check first? is the network link the bottleneck?
I still do not know if the hardware is able to sustain more than a gigabit interface. At least, I have tested servers with 16 processors and they simply cannot sustain the load when you force users enough to fill the gigabit link.
so, my question is: is not there other bottleneck to check first? is the network link the bottleneck?
@Redimido,
if it is office work I agree, if it is a networked database server i'am not so sure yet...., ...Define workload CPU vs. Network load.
From the question we have to assume that the network connectivity is the problem... with as added optional increased throughput.
if it is office work I agree, if it is a networked database server i'am not so sure yet...., ...Define workload CPU vs. Network load.
From the question we have to assume that the network connectivity is the problem... with as added optional increased throughput.
@noci
I understand. my point is you have very well discussed how to add network bandwidth, but the other component will need now to be re-evaluated since it will likely become the new bottleneck on the throughput.
I sustain my say: great thread. I learned a lot with your discussion =^)
I understand. my point is you have very well discussed how to add network bandwidth, but the other component will need now to be re-evaluated since it will likely become the new bottleneck on the throughput.
I sustain my say: great thread. I learned a lot with your discussion =^)
ASKER
That it. 803.2ad in both sides...thanls