LWN: Comments on "Networking and high-frequency trading"

High-frequency trading considered harmful

Wol — Sat, 17 Dec 2022 11:03:04 +0000

Well, the British Empire was so successful because, again, it was business over ideology. Pretty much all of Imperial Expansion was driven by trade - the Pax Britannica ... and it was ended when we lost WWI which - as a trade war - was won by the US (and started as a trade war between Germany and Britain?)

(And no Germany did not actually start WWI. It was started by Austria, and Germany got dragged in as an unwilling participant.)

Cheers,
Wol

High-frequency trading considered harmful

marcH — Sat, 17 Dec 2022 04:11:26 +0000

I recommend "Butler to the World: How Britain Became the Servant of Tycoons, Tax Dodgers, Kleptocrats and Criminals". I didn't read the book, only listened to the author presenting in some podcast and it was enlightening.

The Western world used to colonize the rest of the world by force. Now we got much, much smarter: we let _local_ oligarchs and dictators do the work and then we just take their money because of course they don't trust their own, much more corrupt country to keep it safe. They know it too well.

Business over ideology: isn't that how the Roman Empire lasted 1000 years? Hitler tried to build his empire the other way round and he lasted less than 10 years. Poor thing.

High-frequency trading considered harmful

ssmith32 — Fri, 16 Dec 2022 07:31:15 +0000

If UBS is anything to go by, the Swiss shouldn't be praised for their regulation of financial markets. And the Chinese market is worth whatever Xi wants it to be: c.f. Ma & Ant. It's only as sane as the paramount leader is on that particular day.

HFT and the US market have major issues, but they certainly shouldn't be looking to the Swiss (bankers for *actual* oligarchs), and definitely not the Chinese market.

And even if it's misdirection, I'll take the fixes to the kernel - as long as they pay to make it happen. Even hedgies need to pay up for someone with actual skills now & again ;)

Networking and high-frequency trading

22SAS — Fri, 02 Dec 2022 20:36:39 +0000

Solarflare is the industry standard. Most firms are using one of ef_vi, TCP Direct, or onload.

Networking and high-frequency trading

22SAS — Fri, 02 Dec 2022 20:35:04 +0000

Not all systems use FPGA's, and even FPGA's are constrained to doing certain things. A lot of the stuff is still software based.

Networking and high-frequency trading

22SAS — Fri, 02 Dec 2022 20:32:18 +0000

I work in the industry, at one of the other firms. While this is correct that solutions are mostly using NIC's with vendor supplied kernel bypass libraries, not all system use this solution. Systems that are supposed to be as fast as possible will be using this solution, however, not all applications will require that level of latency. Some applications need to be fast and not that fast, and tuning the kernel's network stack is useful in those cases.

Not all teams will be building sub-microsecond/nanosecond latency systems. The author of that presentation could be on one of those teams.

Networking and high-frequency trading

jepsis — Fri, 02 Dec 2022 07:28:53 +0000

It's amazing WireGuard still has not been fixed. There was a patch for isolated cpus in April but nothing has happened since then and therefore WireGuard still violates kernel behaviour. At minimum WireGuard should make visible in its documentation and Kconfig that it is not compatible with isolcpus.

Networking and high-frequency trading

donbarry — Thu, 01 Dec 2022 17:57:56 +0000

This preposterous way of organizing the world's productive labor really leads to absurdities.

Several companies have been planning undersea cable routes through the Arctic Ocean and the Northwest Passage to connect Asia to Europe, most particularly Japan's stock market with London's. It's not that there's a lack of *bandwidth* that drives these, it's purely latency.

London-Tokyo latency would be reduced by about 60 milliseconds with these cables, if the political knots manage to get untied. The cost? About $1.5 billion.

Networking and high-frequency trading

marcH — Thu, 01 Dec 2022 01:39:37 +0000

Thanks, I was wondering about exactly that.

So in a nutshell they want to have both their cake: highest average performance as possible and eat it: somewhat bounded/predictable latency like a real-time OS would provide.

The word "greedy" keeps coming to me, not sure why :-)

Networking and high-frequency trading

opalmirror — Wed, 30 Nov 2022 00:18:49 +0000

Excellent description. In real-time, the attempt to bound latency adds additional record-keeping and accounting overhead at the lowest levels, which comes at the expense of throughput and CPU and memory cycles. HFT sounds like it optimizes certain paths for highest performance (shortest run time), a very different optimization problem than ensuring high-level algorithms have adequate compute and I/O resources they need to complete algorithms before a particular clock deadline.

Networking and high-frequency trading

marcH — Tue, 29 Nov 2022 15:49:40 +0000

The purpose of "real time" is bounded latency. A scheduler is only one part of the picture.

Networking and high-frequency trading

mathstuf — Mon, 28 Nov 2022 13:46:18 +0000

I don't know that a task defined as "leech money from the system" isn't composed of smaller components that might warrant scheduling against each other.

Networking and high-frequency trading

jem — Mon, 28 Nov 2022 09:19:31 +0000

>Predictable latency, jitter and pretty much everything else screams real-time OS yet "real-time" is never mentioned once... how come?

The term "real-time OS" refers to real-time scheduling, i.e. prioritizing some tasks so that processing is guaranteed to meet some timing constraints. I guess the HFT systems run only one task, so there is no scheduling.

Networking and high-frequency trading

marcH — Sat, 26 Nov 2022 21:53:00 +0000

Forgot to share this great paper, sorry for the noise: https://bravenewgeek.com/everything-you-know-about-latenc...

Networking and high-frequency trading

marcH — Sat, 26 Nov 2022 21:40:39 +0000

Predictable latency, jitter and pretty much everything else screams real-time OS yet "real-time" is never mentioned once... how come?

Disclaimer: I could not find the slides and the "video" link seems not active yet:
https://netdevconf.info/0x16/session.html?The-Anatomy-of-...

There's another thing surprisingly missing: DPDK.

> Being able to use standard Ethernet equipment throughout the network would be worthwhile from a maintenance and cost standpoint as well.

Competing standards and unnecessary duplication of efforts have always been harmed customers and profited vendors but there's only so much "convergence" you can get. Sure, much cheaper and common equipment totally not designed for real time and/or latency is... cheaper. What a surprise?

I'm not familiar with them but there are examples of industries re-using the "good parts" of Ethernet (= the parts that are predictable) and replacing the "bad" ones, here's one:
https://en.wikipedia.org/wiki/Avionics_Full-Duplex_Switch...

There's probably one for cars as well.

High-frequency trading considered harmful

Wol — Sat, 26 Nov 2022 12:36:16 +0000

> > HFT also gives a lot of liquidity to the markets, which is necessary.

> That's complete BS and you know it. The real economy needs liquidity but not measured in thousands of transactions per second.

Liquidity is only needed if you want to use the markets as a gambling den. I hesitate to use the word "properly", but if you're using the markets to store your savings (like pensions, etc) the LAST thing you want is liquidity, and its friend volatility.

You want stodgy, old-fashioned, income stocks that just generate a steady stream of money, and the value of the stocks themselves be damned.

The problem with liquidity, and all this trading, and all that crap, is that stodgy, old-fashioned income stocks become the target of speculation, get taken over by vultures, stripped and the carcasses are left to collapse. All that survives is the name, which gets bought by a company with a far saner market structure (the Swiss come to mind, as do the Chinese), and all we get left with from what was a fine old Anglo-American company is a bunch of rich oligarchs, a load of pension funds in critical care because the company no longer exists, and a load of ex employees with no job and no pension.

And I hate to say it, but when you look at the news coming out of America with Democrats and Republicans talking past each other without hearing, when you look at what's happening over here with middle-class professions like nurses going on strike because they can't afford to live, I think democracy itself could be in serious trouble in the not-too-distant future if we don't do something fast. And a tipping-point climate shock could very easily bring the western house of cards crashing down. Oh well. I think the world, and humanity in small part, will have no trouble surviving. It's just that if Western Civilisation doesn't, most of us in said "Civilisation", won't survive either ...

Cheers,
Wol

High-frequency trading considered harmful

Wol — Fri, 25 Nov 2022 12:23:43 +0000

> A friend working on HFT once described his job as: "stealing your pension". I'm not making this up.

Daytraders - good ones at least - would describe what they do in exactly the same terms.

Let's take an example - Pilkingtons Glass. Was planning to move their listing from London to ?Den Haag. Had they done so, all the big FTSE index linked stocks would have had to dump large holdings, causing the price to seriously implode, and causing lots of damage to the funds. Funds your and my pension is invested in. You make money (short term) in the stock market by predicting what's going to happen, and that was a very easy prediction.

And it's all driven by stupid legislation. A very recent example is the bond market debacle where the Bank Of England was buying bonds like they were going out of fashion. Trying to shore up the price to stop insurance companies (and hence pension funds) going technically bankrupt. If I've promised to give you £1000 in 5 years time, and I have gilts with a face value of £1000 and a redemption date in 5 years time, your money is (barring government default) safe. So why on earth are you worried about ANY gyrations in today's bond values? But because the government says I need £1000 of assets - at today's values! - to cover my 5yr debt, I can be bankrupted today for a debt I can easily repay when it comes due.

Oh - and I nearly lost my pension because of pretty much EXACTLY that silliness ... (and that would have been nearly 20 years ago...)

Cheers,
Wol

High-frequency trading considered harmful

smurf — Fri, 25 Nov 2022 12:06:38 +0000

Well, you won't ever see overlapping bid/ask. The exchange never adds an "overlapping" order to the order book – it executes the trade instead.

> How much of this liquidity provision

IMHO there is no proof whatsoever that HFT adds any liquidity to the market in the first place. On the other hand, AFAIK nobody seriously disagrees with the assertion that it does add a fair amount of volatility and instability.

High-frequency trading considered harmful

paulj — Fri, 25 Nov 2022 10:34:37 +0000

How much of HFT is about seeing overlapping bids and asks, and being fast enough to interpose oneself into the trade and buy the ask and sell to the bid before the actual bidder and asker can do so - thereby 'stealing' a little margin that would otherwise have gone to one of the original parties?

How much of this liquidity provision concerns trades that would have happened at the same - or better - price if the markets involved rate-limited or slot-limited trading?

High-frequency trading considered harmful

marcH — Fri, 25 Nov 2022 08:20:13 +0000

> HFT also gives a lot of liquidity to the markets, which is necessary.

That's complete BS and you know it. The real economy needs liquidity but not measured in thousands of transactions per second.

HFT is pure speculation siphoning money away from the real economy based on the now universal principle "too rich to pay taxes". Even a minuscule per transaction tax (even much smaller than in the real economy) would put an immediate stop to it. HFT is merely exploiting a tax loophole.

A friend working on HFT once described his job as: "stealing your pension". I'm not making this up.

High-frequency trading considered harmful

smurf — Thu, 24 Nov 2022 12:16:48 +0000

Trade fees are ludicrously low. There's no fee or limit on adding nonsense orders to the order books, much less using them to probe what the exchange and/or the other HFTs may or may not be up to.

Also, nobody has yet managed to explain why such liquidity must necessarily occur in microseconds instead of, say, seconds. Or even minutes. "I make money by arbitraging between Frankfurt, London and NYC one millisecond faster than the other guy" (don't see many gals in that space …) is not an endeavor that should be rewarded in a sane economic system IMHO, and I for one won't spend any effort to help those people make even more play money.

Worse: said play money doesn't grow on trees. Ultimately, it's siphoned off from all those who think owning a couple stocks is a good idea but don't have the $$$ to participate in the HFT game. You want to buy for $10 and sell for $20? not if the HFT thing notices. It'll buy for 10.01 and sell for 19.99, leaving $.02 margin for you instead of $10.

Yes I'm exaggerating, and I skipped a couple of levels between the retail stock owner and the HF trader, but WAY not as much as I wish I were.

High-frequency trading considered harmful

ghostbar — Thu, 24 Nov 2022 10:58:36 +0000

There's already a tax on trades, it's called trade fee.

HFT also gives a lot of liquidity to the markets, which is necessary.

Networking and high-frequency trading

farnz — Mon, 21 Nov 2022 10:28:15 +0000

As a bent pipe, it's no good - if you're in London, sending a signal to New York via Starlink, you send the data 550km to a Starlink satellite, which sends it 550km to a ground station near London. It then traverses the fibre route between London and New York, and at the New York end, gets sent 550km into space, then back to the New York station.

This is always going to be slower than cutting out the space hop and using the fibre directly. Where Starlink has the potential to get interesting for latency mavericks is when the inter-satellite links come into use. Then, you'd have the choice between a 5,500 km fibre route at 2/3rd c, or a roughly 7,500 km route via Starlink inter-satellite links at about c; the fibre route is roughly the same as a 8,250 km Starlink route, which gives Starlink an advantage as long as its processing and queueing delays are no worse than 2.5 milliseconds more than the fibre delays.

In practice, the HFT guys will probably pay for low queueing delay, so it all comes down to the processing delays - and I've not seen Starlink publish those details.

Networking and high-frequency trading

caliloo — Sun, 20 Nov 2022 20:08:26 +0000

The variability of starlink links delays might be a new challenge to these guys tho.

Networking and high-frequency trading

caliloo — Sun, 20 Nov 2022 20:06:25 +0000

Starlink still has a shot for over the ocean link. As long as the bent pipe is faster than the light in the glass pipe ...

As for advertisement of the possilibity to these guys : I'd think that starlink does not have to do any adverstisement to make a killing. I'm pretty sure the HFT industry is on permanent alert for new ways to communicate fast.

Networking and high-frequency trading

Paf — Sun, 20 Nov 2022 02:04:46 +0000

Presumably it would compete on all of the *other* routes, though.

High-frequency trading considered harmful

Cyberax — Sat, 19 Nov 2022 19:38:46 +0000

Exchanges don't charge anything for individual trades, though.

High-frequency trading considered harmful

kleptog — Sat, 19 Nov 2022 14:32:13 +0000

Presumably the exchange also wants to be compensated for the trades, so that acts as a de facto tax. You could argue however the exchange isn't charging enough to discourage it (for whatever reason).

Networking and high-frequency trading

mtaht — Sat, 19 Nov 2022 03:22:32 +0000

London to Tokyo via ISL would be very profitable for a starlink-driven approach to trading.

Networking and high-frequency trading

excors — Fri, 18 Nov 2022 09:19:54 +0000

> Starlink right now is not usable for this, because the current generation of satellites work as a "bent pipe", simply beaming traffic back to the nearest ground station.

Even with the inter-satellite laser links and ignoring any processing delays, the satellites are still at 550km altitude (or 340km for future phases if they survive that long), so you're adding ~4 msecs of latency just to go up and down. I suspect that means it'll never be able to compete with ground-based microwave links, when you're willing to spend a lot of money to optimise latency between two fixed endpoints.

Networking and high-frequency trading

Cyberax — Fri, 18 Nov 2022 03:28:47 +0000

> I’m honestly surprised that Starlink doesn’t advertise a faster route for traffic than land lines, and make a killing off HFTing.

There are companies that provide microwave links across the US for this very purpose. Microwaves in air propagate faster than light in glass.

Starlink right now is not usable for this, because the current generation of satellites work as a "bent pipe", simply beaming traffic back to the nearest ground station.

Networking and high-frequency trading

zerolagtime — Fri, 18 Nov 2022 02:57:37 +0000

Michael Lewis’s exposé on HFT revealed an industry that is all about front running. Go fast so that you can raise the price on another buyer putting in a big order, but force them to split their order because of a misused law. Second half is just a few pennies bigger per share traded. If a previous comment is right, and that 50% or more volume is algorithms, just one day netted the industry $10m USD. ($0.02/trade * 1 billion trades in mid Nov 22 * 50%). That’s probably a major lowball.
I’m honestly surprised that Starlink doesn’t advertise a faster route for traffic than land lines, and make a killing off HFTing.
In short, don’t take any sympathy. There’s little incentive for them to cooperate with the Linux community in the form of upstream patches. Not unless there’s something in it for them.

Networking and high-frequency trading

caliloo — Thu, 17 Nov 2022 20:53:09 +0000

Yes, so microseconds of jitter over a trip that takes tens of ms or more is probably fine is my point. Perhaps entering the realm of milliseconds of jitter for a trip that is tens of millis is not.

Networking and high-frequency trading

smurf — Thu, 17 Nov 2022 20:47:20 +0000

> intercontinental arb, which probably doesn’t give a shit about Us as long as it’s not Ms .

You can't get below Ms (as in milliseconds) intercontinentally (OK, so let's ignore Turkey for the moment). Speed of light and all that.

Networking and high-frequency trading

caliloo — Thu, 17 Nov 2022 20:15:38 +0000

HFT is not just about speed of reaction, that is a big component sure (as per definition doh) but the cleverness of your reaction counts as well.

The speeds outlined above for specialised hardware imply a few things: if it takes you 30ns to react that means you can execute a few 100 instructions max. I don’t see the trading strat be very sophisticated there. Plus at these scales speed of light matters. If you want time of travel to be comparable to time of calculations, then only tens or hundreds of feet from the exchange are acceptable.

I’d wager there is a whole world of strategies that require a bit more thought , ie thousands or 10 of thousands of cycles that can accommodate reaction times and time of travel in the us range (which is still only a few miles from the exchange at most). IMHO the interplay between time of reaction and decision complexity allows for a rich world where searching for the right optimization in function of your distance to the exchange is the name of the game.

And there is something to be said for intercontinental arb, which probably doesn’t give a shit about Us as long as it’s not Ms .

That being said, beyond the fun intellectual challenge of being the fastest, I’m also of the opinion that the world would be much better of f with fixed time slot trading , like.. once a day or so… that’d free up a lot of brain cycles to solve the bigger problems of this world.

High-frequency trading considered harmful

Wol — Thu, 17 Nov 2022 20:14:04 +0000

Stagger the tax on how long you hold the stock ... start at 90%, and that drops at 10%/week. And you get charged at the higher rate of trading tax or capital gains tax.

I think UK CGT is about 40%, so trades drop out of the trading tax net at just over a month.

Cheers,
Wol

High-frequency trading considered harmful

smurf — Thu, 17 Nov 2022 17:43:40 +0000

I agree that a tax on trades is a very good idea for multiple reasons, but in order to prevent HFTs you'd need to charge some non-infinitesimal amount for each order, no matter whether it completes or is cancelled.

Also, a tax probably won't address the market instability that's caused by multiple algorithms reacting in microseconds to minimal price changes that humans would simply shrug at.

Also² a tax requires cooperation from politics, which are fed both BS and $$ by the very traders and/or exchanges they're supposed to regulate. The ongoing uphill battles by people who do work towards establishing some sort of transaction tax show that holding your breath until that happens is not a viable survival strategy.

High-frequency trading considered harmful

rgmoore — Thu, 17 Nov 2022 16:54:11 +0000

Another possible solution- one that doesn't require cooperation from the traders or exchanges- is to put a small tax on trades. It would have a minimal effect on people who buy and sell stock because they care about stock as ownership in the company, but it would be a serious deterrent to traders who are just trying to make a profit from market churn.

That said, extreme cases like HFT have often been important test beds for new technologies in Linux. A long time ago, people saw NUMA as an obscure area that was only interesting to people running on exotic big iron. Now NUMA is everywhere, and the effort Linux put into supporting it way back when has paid off. I won't say the stuff people are working on now for HFT is guaranteed to have the same payoff, but it's at least worth paying attention to.

Networking and high-frequency trading

ejr — Thu, 17 Nov 2022 14:30:30 +0000

The talk's numbers are slow even for plain ol' HPC. Years ago.

And the FPGA-based (now called) smart NICs blow the doors off pretty much everything.

High-frequency trading considered harmful

esemwy — Thu, 17 Nov 2022 12:53:58 +0000

Thank you. High speed trading is more about market manipulation than actual commerce. A technological pump and dump, if you will.