Futures/Promises to Lazy Continuations Evolving an Actor Library Based on Lessons Learned from Large-Scale Deployments Benjamin Hindman @benh CppCon 2021prologue ● past life at UC Berkeley, Twitter, Mesosphere/D2iQ ● about a dozen mutexes! (mostly for interfacing with code not written w/ libprocess) ● massive scale (clusters of ~80k physical machines)chapters (1) motivating futures/promises + actors (2) libprocess

int32_t alpha_scale = 0; const int32_t lat = (last_posn.lat + posn->lat) / 2; lon_scale(lat, &alpha_scale, NULL); const uint32_t delta_distance_sc = compute_dist(posn, &last_posn, &alpha_scale, NULL); const int32_t alpha_scale = 0; const int32_t lat = (last_posn.lat + posn->lat) / 2; lon_scale(lat, &alpha_scale, NULL); const uint32_t delta_distance_sc = compute_dist(posn, &last_posn, &alpha_scale, NULL); const int32_t alpha_scale = 0; const int32_t lat = (last_posn.lat + posn->lat) / 2; lon_scale(lat, &alpha_scale, NULL); const uint32_t delta_distance_sc = compute_dist(posn, &last_posn, &alpha_scale, NULL); const

About Us 2 Scott, Principal Engineer at CMU SEI Graph/ML/AI algorithms for large- and small- scale parallel systems. Working on GBTL, a linear algebra-based C++ library for graph analytics.[DISTRIBUTION pointers to custom operators void scale_2(void *out, const void *in) { *(int*)out = 2 * (*(int*)in); } GrB_UnaryOp my_scale_2; GrB_UnaryOp_new(&my_scale_2, scale_2, GrB_INT32, GrB_INT32); built-in types left out of the spec Function pointers (e.g. scale_2) then used in performance-critical inner loops: GrB_apply(C, ..., my_scale_2, A, desc); 57[DISTRIBUTION STATEMENT A] This material has

Point normalized() const { 5 using std::sqrt; 6 const auto scale = 1 / sqrt(x * x + y * y + z * z); 7 return {x * scale, y * scale, z * scale}; 8 } 9 }; Matthias Kretz CppCon ’23 52 GSI Helmholtz Center Point normalized() const { 5 using std::sqrt; 6 const auto scale 7 = 1 / sqrt(x*x + y*y + z*z); 8 return {x * scale, 9 y * scale, 10 z * scale}; 11 } 12 }; 1 void aos(const std::vector>& Point normalized() const { 5 using std::sqrt; 6 const auto scale 7 = 1 / sqrt(x*x + y*y + z*z); 8 return {x * scale, 9 y * scale, 10 z * scale}; 11 } 12 }; 1 void aos(const std::vector>&

errors hard to understand. 15using kilometer_base_unit = bu::make_scaled_unitscale<10, bu::static_rational<3>>>::type; using length_kilometer = kilometer_base_unit::unit_type; using org/z/G3qqn8) (Continued...) 16using kilometer_base_unit = bu::make_scaled_unitscale<10, bu::static_rational<3>>>::type; using length_kilometer = kilometer_base_unit::unit_type; using boost::units::list< boost::units::scaled_base_unitscale<10, boost::units::static_rational<3> > >, boost::units::list

0 码力 | 172 页 | 6.17 MB | 6 月前

3

\ double *B, \ int LDB, \ double beta, \ double *C, \ int LDC)\ { int i,j,k; if (beta != 1.0) scale_c_k18(C,M,N,LDC,beta); if (alpha == 0.||K==0) return; int M4,N8=N&-8,K4; double *a_buffer = (double \ double *B, \ int LDB, \ double beta, \ double *C, \ int LDC)\ { int i,j,k; if (beta != 1.0) scale_c_k18(C,M,N,LDC,beta); if (alpha == 0.||K==0) return; int M4,N8=N&-8,K4; double *a_buffer = (double \ double *B, \ int LDB, \ double beta, \ double *C, \ int LDC)\ { int i,j,k; if (beta != 1.0) scale_c_k18(C,M,N,LDC,beta); if (alpha == 0.||K==0) return; int M4,N8=N&-8,K4; double *a_buffer = (double

behavioral regressions How to use regression testing effectively Establishing a culture of safety at scale8 Aurora Innovation About Aurora Delivering the bene�ts of self- driving technology, safely, quickly productive and safely introduce frequent changes? Implement high-level tests and continuously run them at scale to cover real-world system behaviors with reasonable degree of con�dence.15 Aurora Innovation Developer input. Mismatches against the expected values indicate failure. Tests are dif�cult to maintain, scale, and automate. Testing for Regression Treats a released version of software as baseline. Mismatches

Functionality with CRTP 29 template< typename Derived > struct NumericalFunctions { void scale( double multiplicator ) { Derived& underlying = static_cast(*this); underlying Sensitivity s{ 1.2 }; s.scale( 2.0 ); std::println( std::cout, "s.getValue() = {}", s.getValue() ); }Adding Functionality with C++23 30 struct NumericalFunctions { void scale( this auto&& self, setValue( double v ) { value = v; } double value; }; int main() { Sensitivity s{ 1.2 }; s.scale( 2.0 ); std::println( std::cout, "s.getValue() = {}", s.getValue() ); } No template parameter

Albert EinsteinSo what is there to Rethink?Rethinking: Task Queues Problem #1 - Task Queues Do Not Scale Well: ● Contention: ○ Even the most meticulously designed lock-free queues experience a significant Task Queue Execute Task() Thread Thread PoolRethinking: Task Queues Summary: ● Queues do not scale well: ○ True even for the best lock free implementations ● Does not support prioritization: ○ Available Scalability: ○ Over 40x higher throughput than the fastest MPMC queue at scale ○ Over 100x higher throughput than the average MPMC queue at scale ○ Approximately 1/2N memory requirement (N = number of nodes) Alternative: