Re: [RFC PATCH 0/1] dm-crypt excessive overhead

[Mike Snitzer]

On Wed, Jun 24 2020 at 12:54am -0400,
Damien Le Moal <Damien.LeMoal@xxxxxxx> wrote:

> On 2020/06/24 0:23, Mike Snitzer wrote:
> > On Tue, Jun 23 2020 at 11:07am -0400,
> > Ignat Korchagin <ignat@xxxxxxxxxxxxxx> wrote:
> > 
> >> Do you think it may be better to break it in two flags: one for read
> >> path and one for write? So, depending on the needs and workflow these
> >> could be enabled independently?
> > 
> > If there is a need to split, then sure.  But I think Damien had a hard
> > requirement that writes had to be inlined but that reads didn't _need_
> > to be for his dm-zoned usecase.  Damien may not yet have assessed the
> > performance implications, of not have reads inlined, as much as you
> > have.
> 
> We did do performance testing :)
> The results are mixed and performance differences between inline vs workqueues
> depend on the workload (IO size, IO queue depth and number of drives being used
> mostly). In many cases, inlining everything does really improve performance as
> Ignat reported.
> 
> In our testing, we used hard drives and so focused mostly on throughput rather
> than command latency. The added workqueue context switch overhead and crypto
> work latency compared to typical HDD IO times is small, and significant only if
> the backend storage as short IO times.
> 
> In the case of HDDs, especially for large IO sizes, inlining crypto work does
> not shine as it prevents an efficient use of CPU resources. This is especially
> true with reads on a large system with many drives connected to a single HBA:
> the softirq context decryption work does not lend itself well to using other
> CPUs that did not receive the HBA IRQ signaling command completions. The test
> results clearly show much higher throughputs using dm-crypt as is.
> 
> On the other hand, inlining crypto work significantly improves workloads of
> small random IOs, even for a large number of disks: removing the overhead of
> context switches allows faster completions, allowing sending more requests to
> the drives more quickly, keeping them busy.
> 
> For SMR, the inlining of write requests is *mandatory* to preserve the issuer
> write sequence, but encryption work being done in the issuer context (writes to
> SMR drives can only be O_DIRECT writes), efficient CPU resource usage can be
> achieved by simply using multiple writer thread/processes, working on different
> zones of different disks. This is a very reasonable model for SMR as writes into
> a single zone have to be done under mutual exclusion to ensure sequentiality.
> 
> For reads, SMR drives are essentially exactly the same as regular disks, so
> as-is or inline are both OK. Based on our performance results, allowing the user
> to have the choice of inlining or not reads based on the target workload would
> be great.
> 
> Of note is that zone append writes (emulated in SCSI, native with NVMe) are not
> subject to the sequential write constraint, so they can also be executed either
> inline or asynchronously.
> 
> > So let's see how Damien's work goes and if he trully doesn't need/want
> > reads to be inlined then 2 flags can be created.
> 
> For SMR, I do not need inline reads, but I do want the user to have the
> possibility of using this setup as that can provide better performance for some
> workloads. I think that splitting the inline flag in 2 is exactly what we want:
> 
> 1) For SMR, the write-inline flag can be automatically turned on when the target
> device is created if the backend device used is a host-managed zoned drive (scsi
> or NVMe ZNS). For reads, it would be the user choice, based on the target workload.
> 2) For regular block devices, write-inline only, read-inline only or both would
> be the user choice, to optimize for their target workload.
> 
> With the split into 2 flags, my SMR support patch becomes very simple.

OK, thanks for all the context.  Was a fun read ;)

SO let's run with splitting into 2 flags.  Ignat would you be up to
tweaking your patch to provide that and post a v2?

An added bonus would be to consolidate your 0/1 and 1/1 patch headers,
and add in the additional answers you provided in this thread to help
others understand the patch (mainly some more detail about why tasklet
is used).

Thanks,
Mike