examples/kdigest: add AF_ALG hash example

[ddiss]

When built with CONFIG_CRYPTO_USER_API_HASH enabled, Linux exposes a
socket based API for hashing data. When coupled with uring
IOSQE_IO_LINK, file hashing can be done in quite an efficient manner,
as demonstrated in this link-cp.c based example.

Signed-off-by: David Disseldorp ddiss@suse.de

[ddiss]

v1 was submitted via #335 .

[ddiss]

Changes since previous version:

• add some more crude error checks
• print newline after hash

[ddiss]

@isilence any further thoughts on this patch, or should I drop it?
FWIW, basic benchmark results can be found at #336 (comment)

[ddiss]

@isilence any further thoughts on this patch, or should I drop it?
FWIW, basic benchmark results can be found at #336 (comment)

@axboe feel free to close if there's no interest in this change.

[isilence]

I don't really mind it get merged if it's sane enough, but apparently nobody just have time to take a look. I'd rather assume it to stay until then

[axboe]

@ddiss Were you planning on respinning this change?

[ddiss]

@ddiss Were you planning on respinning this change?

Thanks for the ping. Yes, the respin is still on my todo list, but some other things got in the way. I'm hoping to return to this in a few weeks. I'll close here for now and reopen when ready.

[ddiss]

I finally got a chance to revisit this, so will post an update here, hopefully addressing Pavel's feedback.

[ddiss]

changes since 2021:

• rewrite dispatch logic to submit reads in parallel alongside linked writes I/Os
• use a simple read/write index + state for tracking sequential submission
• preallocate aligned I/O buffers

[axboe]

Thanks for picking this up again! Can you address the CI failures? Looks like it's just basic complaints on things that should be static. I think this would be a good example to add.

[ddiss]

changes to address CI failures:

• drop unused debug macro / statements
• cast off_t for fprintf()

[axboe]

Thanks!

[axboe]

Rather than use IOSQE_IO_LINK, an improvement here may be to use ring provided buffers and bundle send to send data. That should be more efficient, at least on kernels that support both of those features.

[axboe]

73a7003

That adds bundled sends. Didn't test it much, but seems to work.

[axboe]

Ran a quick test on an amd box, and I ranges from ~5x faster at the lower sizes, to being even for 16MB/32MB. Ran your test on it. So whether that helped or not, not sure, but at least it's no longer slower for bigger sizes and it's still considerably faster for sizes that people would actually use.

[ddiss]

Thanks for the follow up changes. FWIW, I did play around with merging buffers for the send, but didn't see much perf improvement for the added complexity. Bundling looks good.

[axboe]

Right, with bundles it's just as easy as using links, at least. But I like having both in the example, as it shows how to use either one.

[ddiss]

In case anyone is interested, below are some rough kdigest ( c90e5a6 ) perf measurements:

kdigest-c90e5a6

  \ FILE
   \SIZE    512 bytes     |       4096       |      65536       |     1048576      |     16777216     |     33554432     |
HASH\__|__________________|__________________|__________________|__________________|__________________|__________________|
   md5 | 0.0010984+-1.44% | 0.0008265+-2.65% | 0.0009556+-1.52% | 0.0024098+-0.35% | 0.0266533+-0.06% | 0.0521402+-0.19% |
  sha1 | 0.0011012+-1.59% | 0.0009430+-2.21% | 0.0009173+-1.89% | 0.0019097+-1.29% | 0.0186466+-0.18% | 0.0361400+-0.13% |
sha224 | 0.0010983+-1.29% | 0.0010970+-1.72% | 0.0010350+-1.32% | 0.0036209+-0.42% | 0.0425834+-0.06% | 0.0841893+-0.06% |
sha256 | 0.0010996+-1.34% | 0.0011159+-1.74% | 0.0010299+-1.09% | 0.0036085+-0.40% | 0.0426466+-0.12% | 0.0840805+-0.03% |
sha384 | 0.0011094+-1.58% | 0.0011204+-1.48% | 0.0009555+-2.96% | 0.0027775+-0.58% | 0.0296736+-0.27% |  0.058271+-0.27% |
sha512 | 0.0010909+-1.71% | 0.0010763+-3.21% | 0.0009746+-1.77% | 0.0027719+-0.74% | 0.0297744+-0.26% |  0.058239+-0.25% |

openssl-3.1.4-3.2

  \ FILE
   \SIZE    512 bytes     |       4096       |      65536       |     1048576      |     16777216     |     33554432     |
HASH\__|__________________|__________________|__________________|__________________|__________________|__________________|
   md5 | 0.0039263+-0.81% | 0.0029834+-1.69% | 0.0030833+-1.28% | 0.0044167+-0.87% | 0.0286969+-0.20% | 0.0536009+-0.16% |
  sha1 | 0.0039302+-1.16% | 0.0029809+-1.62% | 0.0030169+-0.99% | 0.0040051+-1.04% | 0.0220672+-0.29% | 0.0414711+-0.19% |
sha224 | 0.0039211+-0.99% | 0.0039417+-0.95% | 0.0031360+-1.43% | 0.0055392+-0.69% | 0.0408525+-0.18% |  0.078564+-0.20% |
sha256 | 0.0039277+-0.60% | 0.0039653+-0.97% | 0.0031659+-1.26% | 0.0055284+-0.76% | 0.0408774+-0.11% | 0.0788206+-0.10% |
sha384 | 0.0039370+-1.13% | 0.0039494+-0.87% | 0.0030840+-1.26% | 0.0047742+-0.81% |  0.029442+-0.35% |  0.056091+-0.24% |
sha512 | 0.0039456+-1.02% | 0.0039779+-1.09% | 0.0030739+-1.26% | 0.0047586+-0.55% | 0.0294435+-0.20% |  0.056350+-0.31% |

Benchmark script:

for size in $((32 * 1024 * 1024)) $((16 * 1024 * 1024)) $((1024 * 1024)) \
            $((64 * 1024)) $((4 * 1024)) 512; do
    dd if=/dev/urandom of="${size}.data" bs="$size" count=1 || break
    echo "==== hashing file of size $size ===="
    for i in md5 sha1 sha224 sha256 sha384 sha512; do
        # prime cache
        cat "${size}.data" > /dev/null
        perf stat --null -r 5 --table openssl "$i" "${size}.data" >/dev/null 2>openssl.${size}.${i}.perf
        perf stat --null -r 5 --table ~/liburing/examples/kdigest "$i" "${size}.data" >/dev/null 2>kdigest.${size}.${i}.perf
    done
done

Host details:

    Kernel: openSUSE Tumbleweed 6.11.0-1-default
    CPU: Intel(R) Xeon(R) CPU E3-1260L v5 @ 2.90GHz
    Thread(s) per core:   2
    Core(s) per socket:   4
    Socket(s):            1
    RAM: 64GB