24.10.2014 Views

Dynamical HISQ - University of Glasgow

Dynamical HISQ - University of Glasgow

Dynamical HISQ - University of Glasgow

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

<strong>Dynamical</strong> <strong>HISQ</strong><br />

Craig McNeile<br />

c.mcneile@physics.gla.ac.uk<br />

<strong>University</strong> <strong>of</strong> <strong>Glasgow</strong><br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 1


Overview <strong>of</strong> dynamical <strong>HISQ</strong><br />

I report on the various efforts to generate gauge<br />

configurations with <strong>HISQ</strong> sea quarks.<br />

Unfortunately running dynamical simulations<br />

involves code optimization and efficiency.<br />

Sorry!<br />

No results. Sorry! Sorry!<br />

I will try be positive and pretend that the UK will<br />

get computer time for this or (anything). (If time,<br />

end with Iwasaki action).<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 2


Why do dynamical <strong>HISQ</strong>?<br />

From science case for computer time in UK.<br />

Improved scaling from more improved <strong>HISQ</strong> action and<br />

updated radiative corrections to gauge action.<br />

Run simulations with non-degenerate u,d,s and c<br />

quarks.<br />

Include electromagnetism effects. (Check MILC’s<br />

result for m u /m d from 2+1 calculations).<br />

Study Random Matrix Theory (RMT) with <strong>HISQ</strong> given<br />

that the unquenched Asqtad RMT project was<br />

unsuccessful.<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 3


Who is doing what?<br />

Kit and Eduardo’s work discussed later. Who is<br />

writing code.<br />

Person where project<br />

Doug Toussaint Arizona <strong>HISQ</strong> into MILC code<br />

Alexei Bazavov Arizona <strong>HISQ</strong> into MILC code<br />

Carleton DeTar Utah level 3, <strong>HISQ</strong><br />

Tommy Burch Utah level 3, <strong>HISQ</strong><br />

Alan Gray EPCC, UK optimization level 3<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 4


<strong>HISQ</strong> optimization<br />

Ron is a PI for a s<strong>of</strong>tware position in EPCC The grant was<br />

for 1/2 a person (Alan Gray) for 18 months. Alan started at<br />

20 % level in October.<br />

Alan has updated Kit’s code so that it can write configs<br />

in NERSC format. Also more Fortran90.<br />

Alan is studying implementation <strong>of</strong> Asqtad/<strong>HISQ</strong><br />

forces.<br />

Plan for Alan to optimize the level 3 code for<br />

Cambridge machine, and maybe Bluegene/P.<br />

The grant proposal also discussed running and testing.<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 5


<strong>HISQ</strong> and the MILC code<br />

Plaquette test <strong>of</strong> MILC code versus Kit’s code.<br />

Tested Alexei’s MILC inverter against<br />

Eduardo’s on 4 3 8 lattice.<br />

Doug Toussaint used <strong>HISQ</strong> inverter on 1/8 fm<br />

configs by computing the pion spectrum on<br />

an ensemble (m l /m s 0.010/0.050). Good<br />

agreement with HPQCD results.<br />

MILC will write new optimized versions <strong>of</strong> the<br />

<strong>HISQ</strong> routines, as more runs are done.<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 6


RHMC<br />

detM α =<br />

∫<br />

Dφ † Dφe −φ† r 2 (M)φ<br />

where r(x) = x −α/2 (rational approximation).<br />

RHMC is an exact algorithm, also allows larger δt than<br />

R algorithm ( δt ∼ 2/3m l ) so is faster (Kennedy/Clark<br />

claim 7 times better).<br />

MILC are now using the exact RHMC/RHMD algorithm<br />

for generating configurations (talk at ILDG11).<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 7


RHMC improvements<br />

There are a tool kits <strong>of</strong> faster algorithms<br />

Higher order integrators (Omelyan ). In MILC code.<br />

n-th root trick. detM =| detM 1/n | n =<br />

∏ nj=1<br />

∫<br />

Dφ † Dφe −φ† j M −1/n φ j<br />

Mass preconditioning<br />

Chris Richards (Liverpool) used the above algorithms<br />

(implemented by Clark) to get a 30% to 50% speed up with<br />

Asqtad, but needed ∼ 4 months <strong>of</strong> tuning (Lots <strong>of</strong><br />

parameters and trade<strong>of</strong>fs).<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 8


MILC plans<br />

Email from Doug Toussaint. Not started production<br />

running, but deciding this month what they want to do with<br />

<strong>HISQ</strong> (SCIDAC deadline).<br />

Alexei and Doug seeing some issues.<br />

Large jumps in the force when a first-level<br />

smeared matrix is small or when the<br />

phase <strong>of</strong> a unitarized matrix jumps.<br />

Note/email to follow.<br />

Mass dependent Naik term means additional<br />

pseudo-fermion fields are needed (Doug).<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 9


What is level3?<br />

The famous s<strong>of</strong>tware diagram from USQCD.<br />

The level3 library contains the inverter and force<br />

terms optimized for specific platforms. QOPQDP<br />

is the level3 library largely written by DeTar and<br />

Osborn.<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 10


Level3 in practise<br />

Level3 code inside chroma Asqtad inverter.<br />

convert_chroma_to_qdp(in,q_source) ;<br />

convert_chroma_to_qdp(out,psi) ;<br />

qopout = QOP_create_V_from_qdp(out);<br />

qopin = QOP_create_V_from_qdp(in);<br />

QOP_asqtad_invert(&info, fla, &inv_arg, &<br />

mass, qopout,qopin);<br />

QOP_extract_V_to_qdp(out,qopout) ;<br />

convert_qdp_to_chroma(psi,out) ;<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 11


Is level 3 useful?<br />

Machine Library iters Time s<br />

QCDOC Pure chroma 736 46.9<br />

QCDOC single prec. level 3 2932 6.0<br />

QCDOC double prec. level 3 742 2.6<br />

Opteron Pure chroma 1851 8893<br />

Opteron double prec. level 3 1862 4008<br />

Table 1: Performance <strong>of</strong> Asqtad<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 12


Pr<strong>of</strong>iling (Woloshyn and Wong)<br />

100<br />

(a)<br />

cost <strong>of</strong> different components per trajectory<br />

80<br />

percentage<br />

60<br />

40<br />

20<br />

0<br />

8 4 12 4 16 4 8 4 12 4 16 4<br />

ASQTAD<br />

<strong>HISQ</strong><br />

matrix<br />

inversion f (0) f (1) f (2)<br />

f (3) gauge comp. <strong>of</strong><br />

force<br />

eff. links<br />

others<br />

On an Opteron Kit’s code generated a trajectory for a 16 4<br />

lattice in 4 hours. Target <strong>of</strong> 3000 trajectories will take 1.4<br />

years. <strong>Dynamical</strong> <strong>HISQ</strong> – p. 13


What did HPQCD ask for ??<br />

Postdoc and running the MILC code on<br />

Time on Cambridge cluster<br />

Time on Bluegene/P at Darsebury lab in Cheshire.<br />

One rack has been delivered. ETMC twisted mass<br />

HMC code 15% efficiency (after a day’s work by<br />

Carsten Urbach).<br />

3 year project costing £1,000,000.<br />

Amazon.com <strong>of</strong>fer cloud computing. Two cores for 1 hour<br />

with 7 Gbytes <strong>of</strong> memory costs $ 0.4 (20p). (The<br />

Cambridge machine costs 6p per core hour). Any spare<br />

book-tokens could come in useful.....<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 14


Introduction to Iwasaki action<br />

Perhaps 50% <strong>of</strong> lattice QCD computer cycles (JLQCD,<br />

RBC, CP-PACS) use the Iwasaki (square plus<br />

rectangle) action.<br />

A classic book is one that everyone knows about, but<br />

no one has read.<br />

How to do α s n f a 2 corrections (rather than to just annoy<br />

people).<br />

Iwasaki’s paper (1983) “Renormalization Group Analysis <strong>of</strong><br />

Lattice theories and improved lattice action II” not<br />

published (available in PDF on spires) The derivation is<br />

perturbative <strong>Dynamical</strong> <strong>HISQ</strong> – p. 15


The basic idea<br />

Based on RG transformation work by Wilson<br />

e −βS′ (Φ) =<br />

∫<br />

dφe −β(T(Φ,φ)+S(φ))<br />

Fixed point solution has no a m (for any m) or one loop<br />

corrections (section 10.4 <strong>of</strong> DeGrand-DeTar book).<br />

Iwasaki’s block spin transformation on gauge potential.<br />

A (n)<br />

µ (x ′ ) = 1 8<br />

∑<br />

µ (x)<br />

x∈x ′ A (n−1)<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 16


The calculation<br />

Iwasaki looked at effect <strong>of</strong> blocking (n-th level) on one loop<br />

expression for Wilson loops<br />

F (n) (IxJ) =<br />

W (n) (C) = 1 − g 2N2 − 1<br />

N<br />

F (n) (C)<br />

∫<br />

12 12(k)( sin1/2Ik(n) 1<br />

sin 1/2k (n)<br />

D (n)<br />

1<br />

sin1/2Jk (n)<br />

2<br />

sin 1/2k (n)<br />

2<br />

where D 12,12 related to gluon propagator and H(k) is<br />

blocking factor. Also obtained result for ∞ blocking<br />

) 2 H (n) (k)<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 17


Blocking on 1-loop plaquette F(1x1)<br />

Block Wilson T. Symanzik Iwasaki<br />

0 0.5 0.36626 0.21027<br />

1 0.28810 0.25076 0.18826<br />

2 0.21623 0.20599 0.18431<br />

3 0.19446 0.19173<br />

4 0.18865 0.18794<br />

∞ 0.18649 0.18649 0.18649<br />

Iwasaki gauge action wins! (Tadpole Symanzik??)<br />

<strong>Dynamical</strong> <strong>HISQ</strong> – p. 18

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!