Friday, July 29, 2011

NexentaStor 3.1 available now

After a long, and arduous, release cycle, I am pleased to report that NexentaStor 3.1 is available now.

Customers running existing 3.0 installations may upgrade at no cost.

This release includes a number of key features, including some significant improvements for performance and manageability.
Folks using SCSI target mode will probably see the biggest performance boost relative to earlier editions of NexentaStor, especially those folks using NexentaStor to serve up storage to VMware guests -- thanks to the VAAI offload support that is part of this release. And for the record, yes, this release includes the fix the long standing problem with iSCSI timeouts.

For ZFS fans, this release also includes the updates for ZFS version 28. (This does mean that folks upgrading need to be cautious -- their pools will not be automatically updated to ZFS version 28, but if they are manually updated then there will be no way to move those pools back to an older release. That also means that pools should not be updated in HA cluster configurations unless both cluster partners are updated to NexentaStor 3.1 first.) This also means faster snapshot creation and deletion, and the ability to import pools in read only mode or that have encountered a failure of their SLOG device, making some disaster recovery scenarios much less challenging.

Nexenta Core Platform (our general purpose Debian based open core) will be updated shortly to 3.1 as well -- probably sometime in the next week or so. This will be the same core software as supplied with NexentaStor. (We do strongly encourage folks using Nexenta Core Platform for serving up storage to use our commercial NexentaStor product. There is even a free edition for users with up to 18TB of data.)

We are already (and have been for a while actually) hard at work on the next release, which will be based upon illumos, and include a number of other innovative features. Stay tuned for an update.

Thursday, July 28, 2011

Bombproof taskqs

As part of fixing some recent bugs, I integrated the following into illumos:

734 taskq_dispatch_prealloc() desired
943 zio_interrupt ends up calling taskq_dispatch with TQ_SLEEP

The interesting one is the first of these. The interface is actually called taskq_dispatch_ent(), and is private to the "consolidation" (i.e. for use within bundled code only).

What this interface provides for, however, is a way to bomb-proof your taskq dispatches, if you can arrange for the dispatching state structure (taskq_ent_t) to be allocated in advance. This means you never have to worry about the possibility of a dispatch failing due to insufficient resources.

What's even cooler, is that the cost of the dispatching is much cheaper; taskq_dispatch() was the hottest piece of code on a very busy storage server. Now it goes much much faster, because it is just twiddling some linked list pointers and sending a signal to wake up the thread processing the taskq.

More importantly, the fact that we don't ever have to sleep, or have an expensive call into the kernel memory subsystem, has solved some frustrating "stalls" in the storage subsystem.

I encourage developers working with taskqs in illumos to have a look at this interface. If you can use it, you will simplify your code, and shorten your run-paths. Both of which are good things. The only limitation: this interface is not available for dynamic taskqs. (Which makes sense since dynamic taskqs might need to allocate whole threads.)