On Wed, 2019-08-14 at 14:28 +0000, Luse, Paul E wrote:
So I think there's still a feeling amongst most involved in the
that eliminating the memcpy is likely not worth it, especially without
profiling data to prove it. Ben and I were talking about some other much
simpler things that might be worth experimenting with). One example would be
in spdk_nbd_io_recv_internal(), look at how spdk_malloc(), is called for every
IO/ Creating a pre-allocated pool and pulling from there would be a quick
change and may yield some positive results. Again though, profiling will
actually tell you where the most time is being spent and where the best bang
for your buck is in terms of making changes.
From: Mittal, Rishabh [mailto:email@example.com]
Back end device is malloc0 which is a memory device running in the “vhost”
application address space. It is not over NVMe-oF.
I guess that bio pages are already pinned because same buffers are sent to
lower layers to do DMA. Lets say we have written a lightweight ebay block
driver in kernel. This would be the flow
1. SPDK reserve the virtual space and pass it to ebay block driver to do
mmap. This step happens once during startup.
2. For every IO, ebay block driver map buffers to virtual memory and pass a
IO information to SPDK through shared queues.
3. SPDK read it from the shared queue and pass the same virtual address to do
When an I/O is performed in the process initiating the I/O to a file, the data
goes into the OS page cache buffers at a layer far above the bio stack
(somewhere up in VFS). If SPDK were to reserve some memory and hand it off to
your kernel driver, your kernel driver would still need to copy it to that
location out of the page cache buffers. We can't safely share the page cache
buffers with a user space process.
As Paul said, I'm skeptical that the memcpy is significant in the overall
performance you're measuring. I encourage you to go look at some profiling data
and confirm that the memcpy is really showing up. I suspect the overhead is
instead primarily in these spots:
1) Dynamic buffer allocation in the SPDK NBD backend.
As Paul indicated, the NBD target is dynamically allocating memory for each I/O.
The NBD backend wasn't designed to be fast - it was designed to be simple.
Pooling would be a lot faster and is something fairly easy to implement.
2) The way SPDK does the syscalls when it implements the NBD backend.
Again, the code was designed to be simple, not high performance. It simply calls
read() and write() on the socket for each command. There are much higher
performance ways of doing this, they're just more complex to implement.
3) The lack of multi-queue support in NBD
Every I/O is funneled through a single sockpair up to user space. That means
there is locking going on. I believe this is just a limitation of NBD today - it
doesn't plug into the block-mq stuff in the kernel and expose multiple
sockpairs. But someone more knowledgeable on the kernel stack would need to take
Couple of things that I am not really sure in this flow is :- 1. How memory
registration is going to work with RDMA driver.
2. What changes are required in spdk memory management