Alessandro Decina
@alessandrod
Followers
3K
Following
4K
Media
167
Statuses
4K
15
38
135
This was an anza dev node during its leader slots a couple of minutes ago. TPU is at 7% load under current traffic. We have 93% free capacity. TPU is ready for 1M TPS.
29
70
430
Random thoughts now that things have settled a bit (going to regret saying this in 3, 2, 1. ). What happened the past few days is amazing, and I'm still buzzing. We did great, despite what some of the usual larpers are saying (more on that below). Agave ingested traffic at peak.
42
109
402
ethereum bag holders trying every move to divert attention from the fact that their chain does 10 TPS
19
32
273
Did someone say increase bandwidth?. How about a 72% throughput increase by fixing one bug in tokio, with a ~20 lines patch. git diff --stat tokio/src/. 1 file changed, 23 insertions(+), 15 deletions(-)
27
28
176
Most days I wake up and immediately get an aneurysm from someone being dead wrong on the internet. The other day it was from someone claiming that "software scaling has reached its limits". I shared an agave perf profile screenshot, here's how to read it 🧵
19
38
172
@ryanberckmans dude respectfully you have no clue what you're talking about? Like all software solana has a lot of bugs do you want me to tell you about some of those so you have something grounded in reality to tweet about?. We can literally 100x from here just by fixing some extremely dumb.
7
1
148
If you're a guy in your 20s, do open source. You will go into debt.
22
7
127
me looking at the QUIC threads in agave being mostly idle while ingesting 200k TPS
10
5
116
This is one of the upcoming changes in agave I'm most excited about, it's going to make executing txs 30-40% faster.
@alessandrod @WilfredAlmeida_ btw i just landed the pr to remove program cache from account loader on friday sinec that feature just activated on mainet so im writing the filter_executable_program_accounts fix next week
8
15
103
TPU traffic is about 5x what it was before the madness started, finally some transactions, great job everyone keep it up.
8
8
101
we managed to repro, fix coming, put that cron job away.
why do solana rpcs start running slower after two weeks 😭😭😭.
14
6
99
@aeyakovenko @AdrianoFeria @0xMert_ @anza_xyz If we are not faster than firedancer in a year I will quit my job.
11
10
97
New post ✍️. I wrote a little @rustlang tool to intercept SSL/TLS calls and used it to snoop into Zoom. 👀.
4
34
92
PR 1 of 4 out When I'm done, we ingest 5Gbps with 8 threads on a single jemalloc arena with less than 800MB RSS, which is good enough for now. In comparison now we do single digit Mbps on num_cpus() threads on num_cpus() arenas and 10G+ spikes in RSS.
5
13
91
node caught up in 11 mins, _without_ the repair patch @bw_solana recently cooked. Once we get that in too, we fly.
6
7
92
My takeaways from the past few days of profiling banking stage:. - in 2025 we’ll more than 2x block sizes. No doubt in my mind we can easily do it in the validator, not sure about the rest of the infra. - 400ms is an eternity the scheduler would actually work better with shorter.
13
14
84
unbothered. moisturized. happy. in my lane. focused. flourishing. ingesting 1M tps. over quic TPU. over the internet. amsterdam to singapore.
8
10
80
New post! ✍️. A short summary on the work I've been doing to add BPF target support to rustc 🔗. @rustlang.
3
21
76
today we start looking at shred processing. 4x faster insert into blockstore
8
3
78
reminder that none of these people have a clue about what they're talking about
6
8
76
When I started hacking on agave I was horrified when I found we were mmapping everything. It took a ton of work from the accountsdb team and this is where we are now. Slowest is mmap.Mid is no mmap for accounts, mmap for index.Fastest is no mmap for accounts nor index. All this
9
8
74
Finally, let's look at the QUIC TPU threads. These are the threads that used to be _the_ bottleneck at the start of the year: they're what caused the infamous congestion period around March. We've applied a lot of optimizations since then and look at them now: with current mnb
12
9
75
Finally fixed networking latency. Now we can start fixing slot latency.
5
5
72
In the last few months I've been working on RedBPF - a Rust toolkit to work with BPF - that let's you write both kernel and user space code in Rust 🦀🤯 Here's what it looks like:
3
28
70
Solana devex is bad because the people who wrote the sdk were C developers. We need rust developers with good taste for APIs and that know what zero cost abstractions means to write new sdks.
The paradox of solana devrel is that solana culture has always sacrificed devex in the name of performance. Suggestions to improve devex are almost universally rejected by solana’s best developers.
13
5
71
Testing a scheduler fix that should improve local fee markets and slot latency significantly could we airdrop pudgies again or something thanks.
10
4
68
some validators are ingesting >200k TPS and humming along, imagine if block size didn't suck.
TPU traffic is about 5x what it was before the madness started, finally some transactions, great job everyone keep it up.
4
12
68
“tell toranaga-sama that handshake spam takes too much cpu. Thread jitter is high and the tokio runtime cannot keep up with the load. Tell him a new networking protocol is required.”. “The anjin doesn’t know why slot latency is high”
4
7
64
Last hack before BP. A profile of QUIC TPU ingesting 100k txs. Left is upstream quinn, right w/ my fork. Upstream: quinn does 200k allocs - 88.53% of all allocs . Fork: quinn does 1044 allocs - 3.44% of all allocs. Memory usage is fixed, no longer grows linearly w/ throughput.
9
7
63
me: yeah mmap is bad, if you can't precisely control when you page fault perf becomes unpredictable, we must fix this. them: 12 tps has been achieved we must now focus on nuclear winter, total collapse, society in ashes open your eyes. 🥱.
One precision tactical strike on (52.3675,4.9041) in Amsterdam and bye bye Solana. This is what we mean by ETH needs to be in war mode. The ticker is ETH
4
2
64
If you’re building the future of finance and it’s closed source I regret to inform you that you’ve built the past of finance.
2
17
66
literally all I do is IBRL.
increase bandwidth, reduce latency. send tweet.
6
6
61
Some LLM put me in twitter jail for a day for joking about ending @bw_solana so with the extra time I made agave sigverify 2x faster
7
2
61
We've turned the page with 2.0. Wait until you see 2.1. From now on IBRL only.
@jito_labs @jito_sol also s/o to anza core engineers. v2.0.x is really good and made our system much better.
9
2
58
2.0 vs 2.1 vs master. Very satisfying: progress is inevitable
4
6
57
In 2030 you will have . *checks notes* . …4 second blocks?. Is this parody.
6
5
55
This is amazing and we should flat out copy it.
Watch Solana blocks get filled in real time via the Firedancer dashboard:
6
4
56
The biggest cheat code to get better at programming is to do open source, where you have to read other people's code, modify it, and learn to collaborate with people that don't necessarily share your views or priorities. Hire (and generally fund) open source people.
The biggest cheat code to get better at programming is to read more code. Many software interviews make you solve brainteasers, but this doesn't generate the correct signal. The hard part of engineering is being thrown into the ocean of a large codebase and trying not to drown.
3
8
53
I had a zoom with Toly where I didn't even google him beforehand, and left the chat thinking hm unexpectedly smart 🤔.
Who brought you to Solana?.
3
2
52
@0xMert_ the amount of people larping on this app instead of just fixing some bugs I don't know how you deal with them.
2
2
49
@lex_node @pokerchessman @abhi_pingle @trentdotsol Uh no notes on the eth community as a whole getting outplayed - too many people busy writing 10000 words think pieces about things they don't understand. On the solana side people are embracing network extensions that do things that aren't possible on the L1. Not calling them.
8
7
47
I'm so uninterested in the speculation side of crypto that I didn't know what XRP was until last week. This is why I'll be forever poor 😭.
4
0
46
Open source the indexer and we’ll IBRL.
@empirikal @coinbase @brian_armstrong Our indexers can’t keep up… 😭. It’s an us problem, not Solana or a lie.
3
2
46
"software scaling has reached its limits". the limits:. (this is a mnb validator during its 4 leader slots, leaving here without comment)
11
1
45
.@brooksdotsol is going to fix all RPC problems _by EOW_ as his xmas present to us all. He would never disappoint us! Merry christmas brooks🎄.
OH: this is going to fix all rpc problems.
4
4
44
Wow, folks, I just finally tried FD for the first time. You probably know that I have criticized it and its developers in the past, and I must now admit that I was wrong. I mean, truly incredible stuff. Now I understand why some people are saying that it’s unbeatable, probably.
12
3
44
Starting to PR some of the perf fixes from last week. Latency in shred processing from socket recv to replay recv reduced by over 60%
3
3
45
@0xMert_ @ryanberckmans Aaah you're right! I'm usually pretty good at ignoring the nonsense but there's a specific kind of pontificating "non technical" larper that really gets under my skin. Like I just don't get their theory of mind?! Imagine if I started lecturing people about marine biology!? Why!?.
4
0
44
What did you IBRL today? I did a banking stage. Here's the scheduler
5
3
41
Another satisfying chart: this is the time it takes agave to commit back account changes. Predicted a 2x speedup, got a 4x speedup. Oh well.
10
8
39
@aeyakovenko Toly that book shows how to build a k/v store with better perf than the accountsdb in 8 lines of bash go read it.
4
2
40
I spent the whole day learning the rocksdb source code. Verdict: fire all database people across all companies.
2
1
40
Baby numero 2 is out How do you fancy $many_thousands packets latency down to 0 packets latency?.
5
9
39
The core dev meetings with the firedancer team.
1
0
38
I've had a concept of a plan to fix accounts-db for months now.
2
0
37
A couple of months ago I gave a lightning talk at a RustAU meetup about eBPF, profiling and. tar. The video is now up
5
7
34
when we added the off switch, we should have also added a way for me to override the leader schedule to give my node blocks whenever I want. This is such a waste of time I don't have all day.
3
3
37
A transaction trying to get into a block.
0
3
34
Me whenever I profile a new component in agave.
4
3
33
If people didn't cry anytime SOL goes down I wouldn't even know about it. Probably not good for maximizing my bags, but definitely good for my mental health. I recommend never leaving the vscode window. 🧘♂️.
4
0
33
Are you a solana RPC provider? Do you have your custom QUIC TPU client code and the code is open source? If you link it to me I'll review it for you to make sure you're sending as fast as possible.
1
5
32
There's historically been a bug with LTO in the toolchain that's being worked on right now. Once that's fixed, programs become much smaller and deploy much faster ⚡️.
@daglihet @pumatheuma Docs also say that in the section linked. Public one can deploy, just takes a long time. If you're doing production work, it's time to get production infra.
3
1
31
RPCs! I need your help! This PR improves latency by A LOT and reduces request timeouts. I've fixed some methods, I need to fix more. Please share your metrics on which methods are executed the most and take the longest to produce responses. Thanks!.
3
8
32
I don't know anything about other chains maybe algorand is amazing not trying to throw shade, but I'll never forget the face of that candidate I interviewed who used to work on it, and when grilling them on some VM exploit techniques they realized live on zoom that they were.
Stop staking through Coinbase. their higher ups don't seem to give a fuck about Solana and yet they're making millions of dollars off of their massively staked Solana validators. If you're on ramping to Solana use @sphere_labs
1
1
32
> It also requires a PDF protocol spec and research community that's upstream of your chain clients. Same vibes as those movie scenes where they hack into the FBI by showing some minimized js code in Notepad.exe, followed by a popup triumphantly announcing "hacking complete!!!".
There'll be one global backbone - Ethereum. Ethereum is the backbone of the new global financial system of L2s and L1 apps. No other chain will come close. Mert suggested that Sol can pivot to being the backbone. But Solana will never be the backbone. Here's 5 reasons why. Four
5
4
31
so cringe when people don't realize they're out of their depth.
Great to see that Max has left Eth's extended core dev community. Ethereum is for everybody, and Eth has many earnest critics, but Max wasn't one of them. Max has joined Solana. He writes:.> In my first 100 days, I plan on writing a spec for as much of the Solana protocol as I.
4
0
31
if agave could talk.
so basically im fucking stupid.
2
0
30
Next let's look at the threads you probably most care about: the ones that process the user txs that end up in the blocks. At the very start of the first block, all the workers execute transactions in _parallel_. Parallel good. Unfortunately the work quickly becomes
4
2
33
study, lessons.
✓ Squeeze the system you have¹.
3
3
27
I try to stay away from the noise but this is such an obvious difference. we: "damn nothing works we suck we must fix it". them: "nothing works but that's by design because we are very smart". lmao be humble, literally everything still sucks in this space.
Young Max,. The pastures are far greener on this side. On this side, where decentralization is a means and cargo builds are used for heating up one’s living space. On this blockchain, where we engineer first and research second. You belong here, where the pastures are greener.
2
0
28
Today a real estate agent took me to see a place and it's next to this lol. Who is responsible for this?
4
0
26
Could you stoop spiking TPS while I'm asleep and could we get on that zoom call instead to coordinate so I can profile under load thank you.
4
2
25
TIL that people are using IDLs to generate code for multiple languages. Every 10 years someone reinvents COM. Bring back IUnknown.
5
3
25
saturating_add, literally retracing our steps 🥲.
Earlier today, Sui experienced its first network outage due to a bug in congestion control code, which had been recently upgraded to allow for better shared object utilization. Sui contributors quickly deployed a fix restoring normal network activity in 2.5 hours. The rapid,.
4
0
25
You know if they let us implement typed bytecode this conversation wouldn't be happening smh.
$SDK > $IDL.
5
1
24
Uugh it looks like after I discovered the tokio timeout perf issue earlier this year.(.someone fixed it ( and it was.released in tokio 1.38 (june). We didn't pick it up because we're stuck with.older tokio because of. RPC bugs 🤦♂️.
3
1
24
bullish $PENGU it's going to 42 trillions!. (This is the first airdrop I ever got since I literally didn't have to do anything other than stake lol. Strongly recommend more of this.).
1
0
24
The goal is WW3-grade security. The world computer is Ethereum. The ticker is ETH.
1
1
24
My takeaway from that open letter to hyperliquid is that people are rawdogging their funds with random binaries from the docker registry, but I ask you to give me ssh access to your nodes to profile and you go bUt U sTeAl my kEyz 🥺. Pathetic.
3
1
24
actually mine bitcoin don't just read the source like a fkn moron nerd.
You meet yourself on the day you discover crypto. 1) what year is it?.2) what advice do you give yourself?.
4
3
23
@trentdotsol incredible things happen when you don't take a mutex contended in a tight loop by _all_ the threads for each ingested transaction.
2
0
22
Today the Americans have accidentally discovered that skilled visa programs are abused to underpay skilled migrants, and are angry at the exploited migrants and want to be the ones getting underpaid instead (???).
2
2
23
I have just discovered the "remind me later" button on slack. You'll never see a reply from me again. 🫡.
4
0
23
@aeyakovenko The cache is staying, but not getting used for loading accounts anymore. Account loading is getting simplified, a (mis)"optimization" that causes 30-40% overhead by hitting disk querying the account index is getting removed, in favor of the goated account loader cache Hana has.
4
2
22
if I were to summarize the competition
The End of the Solana Golden Age. In our latest podcast, @drakefjustin discusses how @Ethereum's L2s are set to surpass @solana's performance edge in latency and throughput, potentially ending Solana's golden era
2
0
22
Let's start with what's arguably the most important thread: the scheduler. Of the 1.6s the node is leader for (400ms * 4 slots), the scheduler is busy for _less than 200ms_. If the scheduler was close to its limits, you'd see the timeline being all filled with gray. Instead the
2
1
22
Remember: it is always just dumb bugs. This one is extra dumb.
biggest throttles are on the client side 💀.(512 max batch size and 100ms sleep between batches).
2
1
22