Add PGRBENCH_PGRST_NGNIX_LBS option

[Lakshmipathi]

Run two postgrest process behind nginx while using
PGRBENCH_PGRST_NGNIX_LBS option

Signed-off-by: Lakshmipathi Ganapathi [email protected]

[steve-chavez]

Looks good! Even on the t3a.nano I'm getting better results.

I've added a commit to DRY the systemd services(ideally we'd use systemd template units but NixOS doesn't support them yet NixOS/nixpkgs#80933).

---

Not sure if necessary but I'm thinking we should divide the pool size by half and also limit the number of cores a postgREST instance uses with -maxN.

[steve-chavez]

The throughput is noticeable improved with this change.

Starting from m5a.8xlarge the numbers don't improve that much though. Perhaps it makes sense to try with more postgrest services from this instance?

[Lakshmipathi]

| Perhaps it makes sense to try with more postgrest services from this instance?
Yes, correct. We should launch more postgrest depending on no.of cpu core as the instance size gets larger. I'll check and update this PR.

[Lakshmipathi]

@steve-chavez I compared 2-postgrest vs 4-postgrest vs 8-postgrest running behind nginx loading balancer on m5a.4xlarge The results are:

• Running 8-postgrest performs worse than 4-postgrest. And 2-postgrest almost always matches 4-postgrest results except GetSingleEmbed where 4postgrest performs 10% better than 2-postgrest tests. But 2-postgrest produces lesser failure rates than 4-postgrest tests.

So I think we can stick with 2-postgrest for larger nodes.

[Lakshmipathi]

m5a.4xlarge results for 2pgrest:

GETAllEmbed
10, 0.00,169.88375/s
20, 0.00,229.646455/s
30, 0.00,229.746686/s
40, 0.00,229.231026/s
50, 0.00,229.068655/s
60, 0.00,228.791604/s
70, 0.00,228.391011/s
80, 0.00,228.035968/s
90, 0.00,227.370761/s
100, 0.00,227.281861/s
150, 0.00,226.197691/s

GETSingle
10, 0.00,7438.584307/s
20, 0.00,11579.30724/s
30, 0.00,12692.671314/s
40, 0.00,12865.474932/s
50, 0.00,12713.171118/s
60, 0.00,12720.380587/s
70, 0.00,12674.112608/s
80, 0.00,12674.832778/s
90, 0.00,12554.371533/s
100, 0.00,12402.584401/s
150, 0.00,11080.187045/s

GETSingleEmbed
10, 0.00,5471.4133/s
20, 0.00,7686.033566/s
30, 0.00,8495.004228/s
40, 0.00,8597.757732/s
50, 0.00,8467.980147/s
60, 0.00,8214.79728/s
70, 0.00,7970.438457/s
80, 0.00,7690.462343/s
90, 0.00,7470.313378/s
100, 0.00,7335.754678/s
150, 0.00,7148.318773/s

PATCHSingle
10, 100.00,10858.848584/s
20, 100.00,13002.774312/s
30, 100.00,12342.582563/s
40, 100.00,11982.644391/s
50, 100.00,11815.712447/s
60, 100.00,11222.135804/s
70, 100.00,10629.912204/s
80, 100.00,10340.139062/s
90, 100.00,11044.355901/s
100, 100.00,11126.808167/s
150, 100.00,10152.694127/s

POSTBulk
10, 0.00,2148.589839/s
20, 0.00,3374.731514/s
30, 0.00,3774.171711/s
40, 0.00,4671.508998/s
50, 0.00,5191.022437/s
60, 0.00,4792.415842/s
70, 0.00,5029.591023/s
80, 0.00,5055.010742/s
90, 62.56,3979.339248/s
100, 100.00,3080.60703/s
150, 100.00,3505.736743/s

POSTSingle
10, 100.00,4659.941627/s
20, 100.00,14461.154965/s
30, 100.00,14282.000445/s
40, 100.00,14175.212135/s
50, 100.00,14085.749513/s
60, 100.00,13657.977323/s
70, 100.00,13728.898899/s
80, 100.00,13323.034611/s
90, 100.00,13327.493535/s
100, 100.00,13225.353212/s
150, 100.00,10831.574988/s

RPCGETSingle
10, 100.00,10884.519757/s
20, 100.00,13442.455507/s
30, 100.00,12514.133825/s
40, 100.00,12055.88743/s
50, 100.00,10921.116162/s
60, 100.00,11631.847201/s
70, 100.00,11416.375661/s
80, 100.00,11604.109181/s
90, 100.00,11668.606853/s
100, 100.00,11321.700347/s
150, 100.00,11166.950015/s

RPCGETSingleEmbed
10, 100.00,10628.441665/s
20, 100.00,10859.506363/s
30, 100.00,9965.316021/s
40, 100.00,9641.970824/s
50, 100.00,9463.114438/s
60, 100.00,9131.789124/s
70, 100.00,9140.8121/s
80, 100.00,9128.527185/s
90, 100.00,8918.842725/s
100, 100.00,8954.257024/s
150, 100.00,8899.858167/s

RPCSimple
10, 100.00,12825.936503/s
20, 100.00,15180.398109/s
30, 100.00,13870.734118/s
40, 100.00,13252.331772/s
50, 100.00,12473.767051/s
60, 100.00,11948.896705/s
70, 100.00,12434.750048/s
80, 100.00,11988.523623/s
90, 100.00,12125.807341/s
100, 100.00,11988.546133/s
150, 100.00,10921.755359/s

ReadSchemaPATCHSingle
10,X,2382.945716/s
20,X,3299.565402/s
30,X,3282.621257/s
40,X,3926.619464/s
50,X,2235.424783/s
60,X,3279.679749/s
70,X,2862.638494/s
80,X,2256.366125/s
90,X,2454.225369/s
100,X,3471.486376/s
150,X,2554.885865/s

[Lakshmipathi]

m5a.4xlarge results for 4pgrest:

GETAllEmbed
10, 0.00,170.614599/s
20, 0.00,231.922588/s
30, 0.00,232.181214/s
40, 0.00,231.705512/s
50, 0.00,231.468115/s
60, 0.00,231.28597/s
70, 0.00,230.857095/s
80, 0.00,230.552517/s
90, 0.00,230.041414/s
100, 0.00,229.595409/s
150, 0.00,228.560001/s

GETSingle
10, 0.00,7405.60117/s
20, 0.00,11832.799782/s
30, 0.00,12676.037407/s
40, 0.00,12616.486665/s
50, 0.00,12595.870026/s
60, 0.00,12541.565445/s
70, 0.00,12311.928934/s
80, 0.00,12420.406848/s
90, 0.00,12292.117755/s
100, 0.00,12205.565782/s
150, 0.00,12208.238864/s

GETSingleEmbed
10, 0.00,5568.261713/s
20, 0.00,8385.728524/s
30, 0.00,9412.296249/s
40, 0.00,9730.014922/s
50, 0.00,9786.084404/s
60, 0.00,9732.39799/s
70, 0.00,9668.499487/s
80, 0.00,9518.272593/s
90, 0.00,9422.844297/s
100, 0.00,9257.831648/s
150, 0.00,8877.644366/s

PATCHSingle
10, 100.00,9886.625182/s
20, 100.00,12404.82643/s
30, 100.00,12033.25203/s
40, 100.00,12391.010445/s
50, 100.00,12517.494054/s
60, 100.00,12067.571546/s
70, 100.00,12188.260817/s
80, 100.00,11586.453266/s
90, 100.00,11900.199014/s
100, 100.00,13288.359424/s
150, 100.00,14399.458512/s

POSTBulk
10, 0.00,2215.112789/s
20, 0.00,3328.361876/s
30, 0.00,3995.859737/s
40, 0.00,4818.813979/s
50, 0.00,4569.790863/s
60, 18.00,5168.793611/s
70, 100.00,11434.264307/s
80, 100.00,11456.374788/s
90, 100.00,11582.649866/s
100, 100.00,11532.503857/s
150, 100.00,11363.643744/s

POSTSingle
10, 100.00,11529.260161/s
20, 100.00,14935.814373/s
30, 100.00,11577.060617/s
40, 100.00,13328.104777/s
50, 100.00,11715.846538/s
60, 100.00,11698.322062/s
70, 100.00,11086.142935/s
80, 100.00,11772.837252/s
90, 100.00,12963.9979/s
100, 100.00,13589.886795/s
150, 100.00,13609.034938/s

RPCGETSingle
10, 100.00,12001.460356/s
20, 100.00,15727.544059/s
30, 100.00,13586.428121/s
40, 100.00,14505.518048/s
50, 100.00,15092.36333/s
60, 100.00,13372.400677/s
70, 100.00,14954.681154/s
80, 100.00,14965.920158/s
90, 100.00,14620.942568/s
100, 100.00,13203.978571/s
150, 100.00,14076.422781/s

RPCGETSingleEmbed
10, 100.00,9949.740917/s
20, 100.00,12116.51644/s
30, 100.00,13260.539465/s
40, 100.00,12786.511997/s
50, 100.00,12414.98352/s
60, 100.00,11300.577299/s
70, 100.00,12398.197731/s
80, 100.00,12168.631982/s
90, 100.00,11673.412454/s
100, 100.00,12894.695107/s
150, 100.00,12662.8488/s

RPCSimple
10, 100.00,12212.684308/s
20, 100.00,16116.456268/s
30, 100.00,15757.200785/s
40, 100.00,15717.501769/s
50, 100.00,13736.028933/s
60, 100.00,14161.270077/s
70, 100.00,15167.605859/s
80, 100.00,13689.650277/s
90, 100.00,15023.230141/s
100, 100.00,14775.479953/s
150, 100.00,14521.138266/s

ReadSchemaPATCHSingle
10,X,5591.633602/s
20,X,5158.921181/s
30,X,5606.255246/s
40,X,3555.871148/s
50,X,2549.098715/s
60,X,2317.561562/s
70,X,3978.719069/s
80,X,4731.804075/s
90,X,3058.143679/s
100,X,3792.782466/s
150,X,4679.876274/s

[Lakshmipathi]

m5a.4xlarge results for 8pgrest:

GETAllEmbed
10, 0.00,162.102695/s
20, 0.00,167.266801/s
30, 0.00,166.890546/s
40, 0.00,166.581395/s
50, 0.00,166.072316/s
60, 0.00,166.0359/s
70, 0.00,165.999107/s
80, 0.00,165.777881/s
90, 0.00,165.76718/s
100, 0.00,165.625041/s
150, 0.00,164.860827/s

GETSingle
10, 0.00,7400.468126/s
20, 0.00,11287.323241/s
30, 0.00,11789.385666/s
40, 0.00,11781.563705/s
50, 0.00,11814.291927/s
60, 0.00,11780.464721/s
70, 0.00,11673.873235/s
80, 0.00,11623.75242/s
90, 0.00,11577.617937/s
100, 0.00,11481.671537/s
150, 0.00,11483.446519/s

GETSingleEmbed
10, 0.00,5461.947703/s
20, 0.00,7637.721895/s
30, 0.00,8323.575266/s
40, 0.00,8486.658823/s
50, 0.00,8477.813631/s
60, 0.00,8460.544935/s
70, 0.00,8405.09692/s
80, 0.00,8372.7306/s
90, 0.00,8283.465887/s
100, 0.00,8228.478399/s
150, 0.00,7989.330656/s

PATCHSingle
10, 100.00,8720.025475/s
20, 100.00,8609.818363/s
30, 100.00,8629.976643/s
40, 100.00,8548.0264/s
50, 100.00,8539.481544/s
60, 100.00,8853.921616/s
70, 100.00,9958.391867/s
80, 100.00,11470.972233/s
90, 100.00,13597.46222/s
100, 100.00,11176.092348/s
150, 100.00,14405.841887/s

POSTBulk
10, 0.00,2361.390242/s
20, 0.00,3404.679708/s
30, 0.00,4426.262322/s
40, 0.00,4715.12245/s
50, 0.00,4453.947857/s
60, 0.00,4958.312664/s
70, 0.00,5275.624093/s
80, 4.37,5623.595394/s
90, 96.49,3878.725538/s
100, 100.00,3834.870372/s
150, 100.00,3887.341541/s

POSTSingle
10, 100.00,6988.190913/s
20, 100.00,16126.037944/s
30, 100.00,16335.517486/s
40, 100.00,16294.542876/s
50, 100.00,16212.743317/s
60, 100.00,16229.075534/s
70, 100.00,16334.436484/s
80, 100.00,16091.844034/s
90, 100.00,9708.041988/s
100, 100.00,8307.156272/s
150, 100.00,8304.486936/s

RPCGETSingle
10, 100.00,9612.639173/s
20, 100.00,13467.351909/s
30, 100.00,13812.374895/s
40, 100.00,15589.001573/s
50, 100.00,11443.602625/s
60, 100.00,11359.437064/s
70, 100.00,15262.364197/s
80, 100.00,11365.768415/s
90, 100.00,10222.909947/s
100, 100.00,13447.256149/s
150, 100.00,10483.819286/s

RPCGETSingleEmbed
10, 100.00,10858.619536/s
20, 100.00,11254.455464/s
30, 100.00,15352.509429/s
40, 100.00,11529.922041/s
50, 100.00,11388.98909/s
60, 100.00,15291.939105/s
70, 100.00,9154.213994/s
80, 100.00,10002.361932/s
90, 100.00,12601.939431/s
100, 100.00,9849.533201/s
150, 100.00,14778.43507/s

RPCSimple
10, 100.00,10880.706789/s
20, 100.00,11682.984912/s
30, 100.00,15566.607375/s
40, 100.00,11125.029425/s
50, 100.00,9734.018012/s
60, 100.00,13820.618183/s
70, 100.00,9402.196582/s
80, 100.00,14621.130371/s
90, 100.00,11175.906133/s
100, 100.00,11193.814941/s
150, 100.00,14863.798632/s

ReadSchemaPATCHSingle
10,X,2118.955264/s
20,X,2056.975081/s
30,X,4045.384062/s
40,X,4127.390883/s
50,X,3880.531206/s
60,X,2651.908295/s
70,X,2621.500514/s
80,X,3965.07376/s
90,X,4568.707905/s
100,X,5097.690501/s
150,X,5662.486198/s

[steve-chavez]

So I think we can stick with 2-postgrest for larger nodes.

@Lakshmipathi Cool. Then I think this would be good to merge.

[steve-chavez]

m5a.4xlarge results for 4pgrest:

Btw, I also tried with 4 pgrest and my conclusions where the same - 2 pgrest is better.

steve-chavez@c93cca2