Hi Punter11235, I'm interested in putting together the hardware needed to effectively run PioSolver Edge and perform preflop calculations of relative complexity. My current rough plan is to build a desktop PC using 2 Xeon E5-2670 processors, each 8 core with 2.6Ghz. I wanted to get a sense of the speed possible for the max calculations I would need to produce using this setup, so could you roughly estimate the time it would take to produce a strategy for 100BB effective stack with 2 bet and raise sizes (and AI first/last) on all streets including preflop using this setup? Doesn't have to be precise, just want to get a rough sense of how long something like that would take me, and if there is a superior setup I could put together for less than 10k that would result in significant speed gains over my existing plan. I had one other question regarding the program itself-is it possible to perform a GTO calculation from preflop with say 10 flops and then to manually go through each flop editing/locking a range (e.g. one villain's flop Cbet range) and run the calculation again from preflop and have the program respect the postflop locks? I want to see the impact of a player deviating from GTO postflop on preflop play.

Make sure to read my posts about hardware in this thread. Look at last 100-200 posts and ctrl+f cpu/hardware/intel/server

Those are quite decent CPUs.

Again, read my posts in this thread. 16 x 2.6 = 41.6 as it's sandy bridge you need to deduct some also having 16 cores is worse (for scaling) than having less a bit so my hardware score would be about 39.
In comparison non overclocked i7 5820k gets 19.8. This means that Xeon setup will be less than 2x faster than i7 5820k.

Short answer: it's completely unrealistic tree and it's not even close to being realistic.

Long answer: assuming you go easy on sizes in 3 and 4 bet pots you would need about 5TB of RAM and it would take about 2 months to solve (assuming 72 flop subset)

I don't have any more hardware advice than I already provided in this thread. It's better to ask people who are experts at assembling hardware or find some local hobbyist to help you with that.

To be honest I haven't tested postflop locking on preflop trees but it should work.
The reason I haven't tested it is that it's unrealistic to do. You would need to lock many flops an then solve gain - it take a very long time to do both. I definitely wouldn't recommend it.

So first, when you are locking 100% ranges it's a better idea to just remove a 0% branch from the tree or just build a tree without that branch. That will make things faster. In 1.9 version removing branches will be more convenient than it is now (now you need to navigate to a given branch after building the tree and then go to Tools->Solver->Cut currently selected line)
As to locking turns: it should work but it's not very convenient as there are a lot of turns so you would need to click a lot. It's way easier if you only want to do 100% though as you can just remove checking branch from the tree.

Quote:

Short answer: it's completely unrealistic tree and it's not even close to being realistic.

Long answer: assuming you go easy on sizes in 3 and 4 bet pots you would need about 5TB of RAM and it would take about 2 months to solve (assuming 72 flop subset)

----------------------------------

Ok got it, so if a tree of that scale is out of the question, how much time approximately would I be looking at if I were to limit the preflop bet sizes to 2 and the raise sizes to 1, and have only one bet and raise size for all postflop streets (100BB effective/72 flop)? Don't mean to waste your time, just want to get an idea of the calculation time for a realistic preflop tree before I put all this together. If the above is also unrealistic using the hardware specified (2 xeon E5-2670 processors), plz let me know what would be a more realistic target. Thanks!

The idea behind the preflop solver is to simplify postflop play as much as possible to get decent preflop solutions. You can solve postflop trees with elaborate structure later using those ranges and the postflop solver.
It's better to use simplistic postflop play and more flops in a subset for example. You should cut strategic options as much as possible because full preflop trees are humongous.

If you already have a Pio license (basic or pro) you can construct preflop configs and estimate their size yourself.
The answer to your questions depends what exactly you want (how many preflop exits possible etc.). In general if you want to include 2 bet sizes in postflop play and 4-5 or more preflop exits you are looking at terabytes of RAM.

A more realistic target is postflop play with one bet size/one raise size and possibly many preflop exits. For example a tree like this:

https://gyazo.com/64014b3818d473e4cc8db8ac62254c28 on 72 flops and full postflop play with 1 bet size would be 200GB. Elimintating some useless donk bets on the flop OOP (after calling pre) brings it down to ~160GB. You could try using a smaller subset as well.

Notice that this tree has two open sizes. You are looking at similar tree sizes with limp + raise.

Yet another option is to rent your own server or contact PioCloud and asks what he can have available for you.

tldr; is that multiple sizes preflop are humongous and it's not feasible to run trees using them unless you have monstrous hardware available (16core Xeon with 256GB of RAM is very decent but doesn't qualify as monstrous).

Quote: A more realistic target is postflop play with one bet size/one raise size and possibly many preflop exits. For example a tree like this:

https://gyazo.com/64014b3818d473e4cc8db8ac62254c28 on 72 flops and full postflop play with 1 bet size would be 200GB. Elimintating some useless donk bets on the flop OOP (after calling pre) brings it down to ~160GB. You could try using a smaller subset as well.

-------

Ok that sounds doable then using the setup I had in mind. About how long would the provided example tree above then take to calculate using the 2 xeon processor setup?

Probably around 16-20 hours.

I have problems to get my head around a very simple spot.

OOP 88,ATo
IP all Tx, JJ+ and A2s
Boards TsTcTd7s3c
SPR 10

now there is a raising war to the degree that only quads get raised and they block each others quads obv. still there is a reaction(raise AI w quads) to a range that only consists of quads. how can that be?

Is it possible to limit the amount of CPU resources pio uses? Whenever I run sims while trying to do other stuff I get loads of freezing (on the other stuff).

As to the ranges I think it's obvious: both players' range comprise quads only. They both think they have quads if it goes raise-raise-allin (it never goes this way in reality).

As to the % shown on the right (overall percentages). What we show right now is % from perspective of the player. That means % of a given player range or in other words: "how often a player thinks they are taking a given action".

An alternative way would be to show "how often this really happens" which takes card removal from opponents range into account. While we think that what we show now is more useful piece of information the solver provides both and in the next release it will be possible to switch between them in the viewer, like this:

https://gfycat.com/CourteousUnitedAl...hinchillamouse

Yes. Go to Tools->Configuration and change number of threads (on the right) to something lower than number of virtual cores you have (you can experiment with it but something like 4 on a quad i7 or 3 on quad i5 should be fine).

Have you done performance profiling?

I don't understand the question. Did you post in the wrong thread maybe?

hey punter, Im no computer expert and I maybe way off but why cant we use virtual memory (with a SSD) to suplly for the lack of RAM in bigger trees?

You can (there is an option in the solver to ignore memory checks) but it's not recommended because SSDs are way too slow right now so everything will slow down to a crawl. It would also mean a lot of writes (every iteration). That would wear the SSDs off if you keep doing it. Not recommended.

Hi,

I think it would be cool if the main browser screen had a graph of suit distributions in different ranges. Would this be possible?

Is there a way to get a range of hands with weights from PT4 or HM2 into PIO?

Thank you

What would you like to see on such a graph? How often every suit is in range?

Is there any way to import ranges from HM2 or PT4? What's the format?
If the format is not very silly we will consider adding the option to import it.

Yeah, so like the main screen, but instead of hand rankings, you'd have suits (like dd ss hh cc dx sx hx cx).

I am still not sure what kind if information you are after.
You can already see strategies/evs for specific hand combinations, like here:

https://giphy.com/gifs/l46Cph6tRCTBQCwX6

Is there some information you would still like to see which isn't available on this way?

Estimates are always made assuming rainbow unpaired board (the biggest tree possible).
Suited or paired boards are about 1.5x smaller and mono or same rank flops are about 3x smaller. I guess the first board in your script was either mono-suited or had 3 of the same rank.
This is by design because we want the estimation to be the worst case scenario so you are sure the script with many different boards can run.

If you have 32GB of RAM and are building 29GB trees it's highly recommended to ignore memory check after first tree (lower right in script generation window) as Windows is slow to report freed memory.
It's also recommended to build a bit smaller tree as only having 3GB free means that you will run out if you start some things like internet browser or Skype and running out will slow down your computer to a crawl (swap file is very slow).

I mean an aggregation of that information across all hands.

Hello. In the browser, why are the numbers in the "Frequency across runouts:" at the top right and the boxes for each option below it different? The numbers are very similar, so I'm thinking it has something to do with blockers?

Yes, the boxes on the right with total frequencies show a strategy that is: "how often a player thinks they are doing something if they don't know what the opponent's range is".
The frequencies across runouts are shown as real frequencies - "how often it really happens".
The difference between those is in fact card removal.
In the next version it will be possible to change the type of information displayed in those boxes from "strategy" to "real frequency".

Another question.

When choosing the range for "lock selected combos" for node locking, what happens if you choose a fraction of a hand? I tried doing it, but seemed like no node locking happened and it just calculated it as if there were no nodes locked.

I'm wondering because I want to partially lock hands. For example, I want a hand to bet 50% of the time (so that 50% of that combo is locked) but leave the other 50% unlocked (so that it does whatever GTO dictates that it does). Is this possible? If not, seems like it wouldn't be too hard to implement since it would be a minor change to the node locking you already have in place, and other GTO programs have it.

More questions



I got this by clicking the box on the right with the "<".

1. What does clicking the "<" box do?
2. Why are there two "CHECK"s with different values for each 86s on the bottom right?