Introduce --dummy-sig to solana-bench-tps #6575

ryoqun · 2019-10-26T08:46:27Z

Problem

When profiling on the cryptography-unrelated system logic code path for solana-validator to the extreme, solana-bench-tps can be bottle-necked.

Because solana-validator does a heavy-lifting of ed25519 verification by default, perf top is littered with entries related to it. So I first filtered them out by --dev-no-sigverify to concentrate on the transaction processing code for concurrency and scalability in search of potential bottlenecks in there. Then, in turn, solana-bench-tps's ed25519 signing got the next bottleneck!

I need to disable it too to stress solana-validator enough to really expose throughput issues in TPU/TVU. Ultimately, when most of crypto workloads are offloaded to the GPU, the bottlenecks in TPU/TVU will be the next priority.

Solution

Introduce --dummy-sig flag to solana-bench-tps, which is paired with solana-validator's --dev-no-sigverify.

This is still a draft, not final, rather collecting feedback from you! I think this option might be useful, or is there other better way to profile for the mentioned profiling scenario?

Background

I'm trying to profile solana-validator with Linux's perf, already finding some small wins (stay tuned for more PRs!), by the middle course of familiarizing myself with the code base from another viewpoint.

ryoqun · 2019-10-26T08:49:48Z

bench-tps/src/bench.rs

@@ -306,7 +309,11 @@ fn generate_system_txs(
        .par_iter()
        .map(|(from, to)| {
            (
-                system_transaction::transfer(from, &to.pubkey(), 1, *blockhash),
+                if !dummy_sig {


TODO: I'm pretty sure there is still a few places to be branched like this in the solana-bench-tps. Once sanctioned, I'm work on it to finish this.

ryoqun · 2019-10-26T08:50:56Z

sdk/src/signature.rs

@@ -28,6 +28,12 @@ impl Signature {
        Self(GenericArray::clone_from_slice(&signature_slice))
    }

+    pub fn new_rand() -> Self {
+        let mut random_signature_bytes = [0u8; 64];


Magic number!

sdk/src/signature.rs

ryoqun · 2019-10-26T08:53:11Z

sdk/src/system_transaction.rs

    pubkey::Pubkey,
    signature::{Keypair, KeypairUtil},
    system_instruction,
    transaction::Transaction,
 };

-/// Create and sign new SystemInstruction::CreateAccount transaction


Removed this because this comment is too obvious.

sdk/src/system_transaction.rs

ryoqun · 2019-10-26T08:57:18Z

sdk/src/transaction.rs

@@ -96,6 +97,23 @@ impl Transaction {
        Self::new_unsigned(message)
    }

+    pub fn new_signed_instructions_random(


Add test for this.

ryoqun · 2019-10-26T09:04:46Z

sdk/src/system_transaction.rs

+    space: u64,
+    program_id: &Pubkey,
+) -> Transaction {
+    Transaction::new_signed_instructions_random(


API naming: At this layer, we translate to more higher and use-oriented abstract concept from random (implementation details) to dummy (users just want non-verifable fake signatures cheaply).

codecov · 2019-10-26T09:22:18Z

Codecov Report

Merging #6575 into master will decrease coverage by <.1%.
The diff coverage is 48.1%.

@@           Coverage Diff            @@
##           master   #6575     +/-   ##
========================================
- Coverage    81.8%   81.8%   -0.1%     
========================================
  Files         244     241      -3     
  Lines       51902   51917     +15     
========================================
- Hits        42504   42470     -34     
- Misses       9398    9447     +49

ryoqun · 2019-10-29T00:48:40Z

@mvines I elaborated the description a bit! Could you take a look at it in spare time?

stale · 2019-11-28T01:39:14Z