Tuesday, November 9, 2010

Notes from Solaris Summit, LISA 2010

There was a lot of good recapitulation on features that has been available in OpenSolaris that will be part of Solaris 11 as well as some new things.

Fist of all, Solaris 11 Express is coming any day now, they mentioned weeks and days as the timeframe for the release which I suspect will be build snv_151a (or possible another respin). Given the internal tag of the release 2010.11 they should at least be aiming for a November release.

There was a lot of focus on the integration of zones with different core components of Solaris, IPS was discussed in regard to zones and updates and the synergies of integrating zones with crossbow. In the future (post S11 Express) there will also be possible to control storage inside of zones in a good way, if allowed utilities like format and newfs should be usable on devices dedicated to the zone. The storage focus for also include Zones as NFS servers.

Crossbow will be optimized for NUMA localities, a new framework will be available, NUMA I/O and will at least be used by infiniband besides the networking stack.

Solaris 11 will be able to handle the CPU resources dedicated for a zone even for the network utilization. With crossbow threads can be bound to virtual interfaces and these can be kept in sync with the ones dedicated for the zone making the zone an even more isolated environment.

The rewrite of the virtual memory subsystem (VM2.0) is alive and will be delivered incremental in the Solaris 11 lifetime. This should pave way for thinks like virtualized memory which can be dedicated to zones (Really dedicated RAM and swap to the zone, not the capping we have today) as well as power management of memory. A description of VM 2.0 by Blake A. Jones:

"VM2 is a project to redesign the Solaris virtual memory system around modern computer architectures. The core of the current VM system was designed in 1985-86, when Sun's large computers had 4 megabytes of RAM, one simple CPU with a simple MMU, very few disks, and no NUMA or power management. A couple decades of Moore's Law and many many billions of dollars of hardware development later, things look a bit different. Obviously the software has evolved to deal with these hardware changes, and it's a testament to the original design that it's performed for this long. But the VM system has developed a reputation for being hard to understand, as 20 years' of accreted development will tend to do, and having most of the VM interfaces operate on lists of small fixed-size pages has made it hard to do more significant innovation."

There also was some ZFS news, but that has been covered by this blog before: ZFS crypto, RAID-Z/mirror hybrid allocator and some stuff available in the external source before it closed. One thing new was some numbers on performance impact of crypto and speed enhancement with the new raidz allocator:

"Actual costs for ZFS encryption is 7% for random I/O,
and 3% for sequential I/O

ZFS RAID-Z mirror allocator - preliminary data is that
some workloads are 2 to 4 times faster for things like
directory searching"

There is much more interesting stuff in the videos and presentation slides, both are still available here:

ZFS Crypto integrated
Solaris 11 Express 2010.11 and ZFS
Zones should be able to be NFS servers

No comments: