Guest swap obsolete? ESX and serious memory overcommitment.

Hello!

I would like to hear your considerations regarding getting rid of swapping on the guest OS by doing serious overcommitment of memory on ESX(i) server.

Here is my thoughts:

Wouldn't it be more flexible and efficient to NOT have any swap space on the guest and just allocate more vMemory for the guests?

Considder this imaginary system:

1. ESX host with 8GB of pMemory
2. 4 Guests, each with 8GB of vMemory and NO swapping area for any of them.
3. All 4 guests have supported guest Operating Systems with VMTools enabled (balloon driver).

Will I then be right in the following?:

Pros:

• Because the guests do not each handle their own swapping it can be handled centrally and more effectively by then ESX host. Because old pages are not relative to other pages in the same guest but to all other pages of all other guests on the host.
• Swapping could be much faster if the ESX swap areas was saved on super fast SSD's instead of guest swap which is often HDD's or SAN (swapping over network should be slower than on local SSD disks). Also due to ESXs support for memory sharing, compression, ballooning and  other advanced memory saving capabilities, memory is more effectivly used. It is much better to have 2 guest pages on one shared host memory page, than one guest memory page in memory and the other guest page swapped out to slow HHDs due to tight memory (right?).
• A more clean and simple setup in that swap is more a 'hardware' thing and is located on the local ESX host and data disks is more application orientated on located on SAN (or where all the guests are located).
• Much better memory utilization can be archived because of ESX's support for memory reserves, shares and limits.

Cons:

• Snapshots takes longer and more space to be saved because the snapshot files matches the size of the vMemory of the guest.
• Takes up alog more disk space for storing those swap areas on the ESX host (4*8GB in this example).
• Will it make Live Migration slower?
• Requires better host hardware. Hardware that do not block too much of the other guests just becase on of the guests have alot of page faults.
• If a guest is running out of memory some guest operating systems support adding additional swap area online without rebooting. By having no swap on the guest a reboot (and increase of vMemory) needs to be done. Even though a temporary client based swap area could be added while waiting for the next maintanence reboot of the guest
• Operating systems like Linux and Solaris fill all its unused memory with filesystem cache. By having more memory instead of cache the guest do not know the difference and cache very much filesystem (meta)data. But is this really a Pro or a Con? I mean ESX just keeps the most used pages and if page with filesystem meta data is heavily used then that is a good thing. Maybe this is really a Pro?
• Will the baloon driver work correctly without any guest swap area?

I know the subject of memory overmommitment has been up before but also in the context of totally getting rid of client swap area?

I would love to hear your considderations, other pros/cons and if I am right and wrong in any of mine?

Best reagards,

- Morten Green Hermansen