Better serialization of binary values and singletons #844

iliailia · 2024-12-19T16:27:18Z

Instead of sending single values as VecRingT, we serialize them as individual elements. This is especially beneficial when doing bitwise comparisons, where sending a single bit required 5 bytes of serialized data. This change reduces it to 2. For u64, the size is reduced from 20 to 10 bytes.
In addition, this PR replaces bincode with bitcode, which is about 15% more compact for the NetworkValue enum. The total gain is about 37%.

Communication cost per party to insert 1 vector into a database of size 1 (i.e., one call to eval_distance and one call to less_than) is given below.

Before:

Stats for party Identity("alice"):
Bytes sent: 1389
Bytes received: 1389
Stats for party Identity("bob"):
Bytes sent: 1389
Bytes received: 1437
Stats for party Identity("charlie"):
Bytes sent: 1417
Bytes received: 1369

After:

Stats for party Identity("alice"):
Bytes sent: 885
Bytes received: 885
Stats for party Identity("bob"):
Bytes sent: 885
Bytes received: 907
Stats for party Identity("charlie"):
Bytes sent: 896
Bytes received: 874

bitcode is slower than bincode, which is reflected in a 5-25% increase in running time for both insertion and search.

gr_ready_made_hnsw/gr-big-hnsw-insertions/1
                        time:   [2.3032 ms 2.3137 ms 2.3299 ms]
                        change: [+6.1174% +8.1775% +10.549%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/1
                        time:   [3.3832 ms 3.4097 ms 3.4656 ms]
                        change: [+7.6982% +9.7017% +11.805%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/10
                        time:   [85.330 ms 86.775 ms 88.262 ms]
                        change: [+2.7516% +4.7590% +6.7136%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/10
                        time:   [53.284 ms 53.762 ms 54.214 ms]
                        change: [+6.1638% +10.074% +13.611%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/100
                        time:   [1.2600 s 1.2681 s 1.2769 s]
                        change: [+1.9064% +4.8861% +7.8922%] (p = 0.01 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/100
                        time:   [775.21 ms 786.06 ms 797.46 ms]
                        change: [+10.597% +15.231% +19.069%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/1000
                        time:   [3.4237 s 3.4414 s 3.4590 s]
                        change: [+6.5509% +8.5298% +10.339%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/1000
                        time:   [2.0689 s 2.1586 s 2.2907 s]
                        change: [+10.130% +15.345% +23.125%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/10000
                        time:   [6.9597 s 7.1282 s 7.3122 s]
                        change: [+14.125% +16.964% +20.189%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/10000
                        time:   [3.7279 s 3.8948 s 4.0742 s]
                        change: [+13.594% +19.064% +25.207%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/100000
                        time:   [9.0077 s 9.4853 s 9.9697 s]
                        change: [+19.589% +25.218% +31.718%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/100000
                        time:   [4.6592 s 4.8944 s 5.1631 s]
                        change: [+8.8328% +14.471% +20.709%] (p = 0.00 < 0.05)
                        Performance has regressed.

…many

iris-mpc-cpu/src/network/value.rs

…many

iris-mpc-cpu/src/network/value.rs

iliailia · 2025-01-08T15:44:29Z

Guys, sorry for changing this PR almost entirely. I added a custom serialization that brought the performance back and even gained a bit more. The current benchmarking results are the following

Communication:

Stats for party Identity("bob"):
Rounds: 53
Bytes sent: 845
Bytes received: 875
Stats for party Identity("charlie"):
Rounds: 53
Bytes sent: 862
Bytes received: 832
Stats for party Identity("alice"):
Rounds: 53
Bytes sent: 845
Bytes received: 845

Running time (in comparison to the one above):

gr_ready_made_hnsw/gr-big-hnsw-insertions/1
                        time:   [2.0494 ms 2.0767 ms 2.1024 ms]
gr_ready_made_hnsw/gr-big-hnsw-searches/1
                        time:   [3.0304 ms 3.0676 ms 3.0967 ms]
gr_ready_made_hnsw/gr-big-hnsw-insertions/10
                        time:   [74.011 ms 74.983 ms 76.775 ms]
gr_ready_made_hnsw/gr-big-hnsw-searches/10
                        time:   [45.616 ms 46.257 ms 47.247 ms]
gr_ready_made_hnsw/gr-big-hnsw-insertions/100
                        time:   [1.0943 s 1.1229 s 1.1492 s]
                        change: [-25.943% -23.699% -21.898%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/100
                        time:   [662.72 ms 687.23 ms 712.31 ms]
                        change: [-26.357% -23.826% -20.564%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/1000
                        time:   [2.9798 s 3.0148 s 3.0564 s]
                        change: [-25.662% -24.840% -23.886%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/1000
                        time:   [1.8064 s 1.8388 s 1.8704 s]
                        change: [-22.091% -20.095% -18.120%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/10000
                        time:   [5.6744 s 5.7518 s 5.8269 s]
                        change: [-23.766% -22.686% -21.621%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/10000
                        time:   [3.0832 s 3.1271 s 3.1760 s]
                        change: [-25.487% -24.297% -23.200%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/100000
                        time:   [6.7998 s 6.8641 s 6.9288 s]
                        change: [-26.099% -24.969% -23.805%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/100000
                        time:   [3.4343 s 3.5064 s 3.5838 s]
                        change: [-31.274% -29.707% -27.821%] (p = 0.00 < 0.05)

dkales · 2025-01-08T15:45:11Z

Guys, sorry for changing this PR almost entirely. I added a custom serialization that brought the performance back and even gained a bit more. The current benchmarking results are the following

Communication:

Stats for party Identity("bob"):
Rounds: 53
Bytes sent: 845
Bytes received: 875
Identity("charlie") Open bin bytes: 1
Stats for party Identity("charlie"):
Rounds: 53
Bytes sent: 862
Bytes received: 832
Stats for party Identity("alice"):
Rounds: 53
Bytes sent: 845
Bytes received: 845

Running time (in comparison to the one above):

gr_ready_made_hnsw/gr-big-hnsw-insertions/1
                        time:   [2.0494 ms 2.0767 ms 2.1024 ms]
gr_ready_made_hnsw/gr-big-hnsw-searches/1
                        time:   [3.0304 ms 3.0676 ms 3.0967 ms]
gr_ready_made_hnsw/gr-big-hnsw-insertions/10
                        time:   [74.011 ms 74.983 ms 76.775 ms]
gr_ready_made_hnsw/gr-big-hnsw-searches/10
                        time:   [45.616 ms 46.257 ms 47.247 ms]
gr_ready_made_hnsw/gr-big-hnsw-insertions/100
                        time:   [1.0943 s 1.1229 s 1.1492 s]
                        change: [-25.943% -23.699% -21.898%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/100
                        time:   [662.72 ms 687.23 ms 712.31 ms]
                        change: [-26.357% -23.826% -20.564%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/1000
                        time:   [2.9798 s 3.0148 s 3.0564 s]
                        change: [-25.662% -24.840% -23.886%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/1000
                        time:   [1.8064 s 1.8388 s 1.8704 s]
                        change: [-22.091% -20.095% -18.120%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/10000
                        time:   [5.6744 s 5.7518 s 5.8269 s]
                        change: [-23.766% -22.686% -21.621%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/10000
                        time:   [3.0832 s 3.1271 s 3.1760 s]
                        change: [-25.487% -24.297% -23.200%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-insertions/100000
                        time:   [6.7998 s 6.8641 s 6.9288 s]
                        change: [-26.099% -24.969% -23.805%] (p = 0.00 < 0.05)
gr_ready_made_hnsw/gr-big-hnsw-searches/100000
                        time:   [3.4343 s 3.5064 s 3.5838 s]
                        change: [-31.274% -29.707% -27.821%] (p = 0.00 < 0.05)

nice, have a few more suggestions that might help a bit more with perf

bgillesp

Nice! See minor comments, but looks good to me if Daniel has finished suggesting optimizations.

iris-mpc-cpu/Cargo.toml

iris-mpc-cpu/src/network/value.rs

bgillesp · 2025-01-08T20:26:08Z

iris-mpc-cpu/src/network/value.rs

@@ -18,20 +17,250 @@ pub enum NetworkValue {
 }

 impl NetworkValue {
+    fn get_descriptor_byte(&self) -> u8 {
+        match self {
+            NetworkValue::PrfKey(_) => 0x01,


Do these descriptor bytes start at 1 because the network byte stream looks for 0 to indicate EOF or something similar?

I was thinking to leave it for some special message, e.g., "abort"

iris-mpc-cpu/src/network/value.rs

iliailia added 4 commits December 16, 2024 10:30

Reduce messages in and_many

ae80a94

Merge remote-tracking branch 'origin/main' into ilia/reduce-comm-and-…

189f23f

…many

Serialize with bitcode

f43c138

Remove Ring16 and Ring32 network values

63db5fd

iliailia requested review from bgillesp and dkales December 19, 2024 16:32

dkales approved these changes Jan 8, 2025

View reviewed changes

iris-mpc-cpu/src/network/value.rs Show resolved Hide resolved

iliailia added 2 commits January 8, 2025 13:36

Merge remote-tracking branch 'origin/main' into ilia/reduce-comm-and-…

365103c

…many

Custom serialization

7ead648

dkales reviewed Jan 8, 2025

View reviewed changes

iris-mpc-cpu/src/network/value.rs Outdated Show resolved Hide resolved

iris-mpc-cpu/src/network/value.rs Outdated Show resolved Hide resolved

iris-mpc-cpu/src/network/value.rs Outdated Show resolved Hide resolved

Better vector allocation

69bec46

bgillesp approved these changes Jan 8, 2025

View reviewed changes

dkales reviewed Jan 9, 2025

View reviewed changes

iris-mpc-cpu/src/network/value.rs Outdated Show resolved Hide resolved

Address comments

556ed13

iliailia merged commit d31df96 into main Jan 9, 2025
13 checks passed

iliailia deleted the ilia/reduce-comm-and-many branch January 9, 2025 14:38

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Better serialization of binary values and singletons #844

Better serialization of binary values and singletons #844

iliailia commented Dec 19, 2024 •

edited

Loading

iliailia commented Jan 8, 2025 •

edited

Loading

dkales commented Jan 8, 2025

bgillesp left a comment

bgillesp Jan 8, 2025

iliailia Jan 9, 2025

Better serialization of binary values and singletons #844

Better serialization of binary values and singletons #844

Conversation

iliailia commented Dec 19, 2024 • edited Loading

iliailia commented Jan 8, 2025 • edited Loading

dkales commented Jan 8, 2025

bgillesp left a comment

Choose a reason for hiding this comment

bgillesp Jan 8, 2025

Choose a reason for hiding this comment

iliailia Jan 9, 2025

Choose a reason for hiding this comment

iliailia commented Dec 19, 2024 •

edited

Loading

iliailia commented Jan 8, 2025 •

edited

Loading