Posts Tagged ‘cost’

Limiting the number of concurrent storage vMotions

June 6, 2013

vmw-dgrm-vsphr-087b-diagram1VMware vCenter allows several concurrent storage vMotions on a datastore. But it can negatively impact your production environment, by hammering your underlying storage. If you want to migrate several virtual machines to another datastore, it’s much safer to do that one by one. But it’s too much manual work.

There is a simple way to limit the number of concurrent storage vMotions by configuring vCenter advanced settings. There are a group of resource management parameters for network, host and datastore limits which apply to vMotion and Storage vMotion. They are called limits and costs. For ESXi 4.1 default datastore limit for migration with Storage vMotion is 128. And datastore resource cost for Storage vMotion is 16 (defaults for other versions of ESXi can be found here: Limits on Simultaneous Migrations). It basically means that 8 concurrent storage vMotions is allowed for each datastore. So to allow only one storage vMotion at a time you can either change the limit to 16 or cost to 128.

Lets say we choose to change the cost to 128. There are two ways of doing it. The first one is to edit vCenter vpxd.cfg file and add the following stanza between <vpxd></vpxd> tags:

<ResourceManager>
<CostPerEsx41SVmotion>128</CostPerEsx41SVmotion>
</ResourceManager>

The second simpler one way is to edit vCenter -> Administration -> vCenter Server Settings -> Advanced Settings and add config.vpxd.ResourceManager.CostPerEsx41SVmotion key with value equal to 128. You will probably need to reboot vCenter after that.

There is one moment, however. If you migrate VMs from say 3 source datastores to 1 destination, then 3 concurrent storage vMotion will kick off. I do not know what is the reason for that, but that’s what I found from the practice.

Advertisements

Spanning Tree Protocol Overview

July 16, 2012

When it comes to switching it is recommended to understand how STP works. STP was developed to prevent loops. For example, you connect 3 switches in a ring, some host sends a broadcast packet. Since broadcast packet is flooded to all ports (forget about VLANs for a moment) it will travel several times around the ring until its TTL is equal to 0. This situation will never happen if you work on Cisco switches. They have STP enabled by default. Some low-budget switches do not support STP at all.

To prevent loops STP disables some ports or in other words put them in a blocking state. Ports that are left to forward traffic are in a forwarding state. To exchange STP information switches use Bridge Protocol Data Units (BPDU). They contain three main fields: root switch ID, sender switch ID and cost to reach the root. ID is almost random and are based on priorities and MACs. Cost depends on link speed. 100Mb port’s priority equals to 19, 1Gb is 4, etc.

STP starts from electing a root switch. All switches exchange their IDs and switch with the lowest ID becomes a root switch. As stated above root switch is almost a random choice, but you can manually assign priority if needed. Then spanning tree algorithm (STA) searches for root ports (RP) and designated ports (DP). RP is a port with the shortest path to the root switch. Shortest path is founded based on link weights and if they are equal on switch IDs. DP is a port with the lowest cost to the root on that Ethernet segment. Ethernet segment here is a collision domain, which in its turn in switched network is simply an Ethernet link between two switches. Basically, that means that you will have one shortest path from each non-root switch to the root switch. On one side of each link will be a RP and on the other a DP port. All non-shortest paths will have DP on one side and non-DP non-RP  (blocked) port on the other side. Traffic will not traverse through this port to prevent loops.

You may ask, what’s the point of such distinction between DP and RP in this concept if the only thing that matters is the shortest path. Even though RP and DP lies on the shortest path to the root, just from the opposite sides, there is one significant distinction between them. DP is the port from which Hello BPDUs are continuously sent. Hello BPDU simply indicates that link between switches is working and contains information which allows switch on the other side of the link to find the new shortest path to the root in case an old link brakes. Another difference is that DPs exist not only on root paths, but on each of the Ethernet links.

Along with STP, there is a RSTP, which stands for Rapid Spanning Tree Protocol. The reason for RSTP is that STP converges slowly. Convergence is a process which happens when network topology changes and switches need to reevaluate port statuses (blocking/forwarding). STP converges for approximately 50 seconds. RSTP convergence time is 1 to 10 seconds.

STP and RSTP have several implementations. Cisco by default uses PVST+ (or simply PVST) which is an abbrevation for Per-VLAN Spanning Tree Plus, instead o pure IEEE’s STP. PVST creates one STP topology per VLAN. Instead of using one link for all VLANs and block all other links, you can use first link for even VLANs and second for odd. PVST allows you to do that. Cisco’s implementation of RSTP is called PVRST (Per-VLAN Rapid Spanning Tree) or RPVST (Rapid Per-VLAN Spanning Tree). There is an IEEE implementation of protocol similar to PVRST. It’s called MIST – Multiple Instances of Spanning Trees. MIST is an implementation of RSTP. MIST’s difference from PVRST is that it doesn’t create separate STP for each VLAN as PVRST does by design, but lets you create one STP for multiple VLANs.