Update figures

Author: Cecca

.gitignore

@@ -1,4 +1,4 @@
-danny-results.sqlite 
+*.sqlite 
 /env
 .vscode/
 .smt

analysis/plot_k_dep.R

@@ -32,7 +32,8 @@ plot_counters <- function(data, t, ylabels = FALSE) {
     p_candidates <- ggplot(data, aes(
         x = k,
         y = fraction_candidates,
-        color = algorithm
+        color = algorithm,
+        shape = algorithm
     )) +
         geom_line(size = 0.4, alpha = 0.5, position = position_dodge(width = 0.5)) +
         geom_point(size = 0.5, position = position_dodge(width = 0.5)) +
@@ -75,7 +76,8 @@ plot_counters <- function(data, t, ylabels = FALSE) {
     p_load <- ggplot(data, aes(
         x = k,
         y = Load,
-        color = algorithm
+        color = algorithm,
+        shape = algorithm
     )) +
         geom_line(size = 0.4, alpha = 0.5, position = position_dodge(width = 0.5)) +
         geom_point(size = 0.5, position = position_dodge(width = 0.5)) +

analysis/plot_scalability.R

@@ -3,68 +3,106 @@ source("tables.R")
 source("plots.R")
 
 plotdata <- table_scalability() %>%
-    filter(threshold == 0.5) %>%
-    filter(sketch_bits == 256, k2 %in% c(0, 6)) %>%
+    filter(dataset == "Glove") %>%
     group_by(algorithm, dataset, workers, k, k2, sketch_bits) %>%
     summarise(total_time = mean(total_time)) %>%
     mutate(total_time = set_units(total_time, "s") %>% drop_units()) %>%
     ungroup() %>%
-    select(algorithm, dataset, workers, k, k2, sketch_bits, total_time)
-
-fast_params <- plotdata %>%
-    group_by(algorithm, dataset, workers) %>%
-    slice_min(total_time) %>%
-    ungroup() %>%
-    select(algorithm, dataset, workers, k, k2, sketch_bits)
-
-best_40_params <- plotdata %>%
-    filter(workers == 40) %>%
+    select(algorithm, dataset, workers, k, k2, sketch_bits, total_time) %>% 
     group_by(algorithm, dataset, workers) %>%
-    slice_min(total_time) %>%
-    ungroup() %>%
-    select(algorithm, dataset, k, k2, sketch_bits)
-
-with_best_40 <- semi_join(plotdata, best_40_params) %>%
-    mutate(conf = "best40")
+    slice_min(total_time) %>% 
+    ungroup()
 
-fast_runs <- semi_join(plotdata, fast_params) %>%
-    mutate(conf = "bestWorker")
-
-plotdata <- bind_rows(
-    with_best_40,
-    fast_runs
-) %>%
-select(-k, -k2, -sketch_bits) %>%
-pivot_wider(names_from=conf, values_from=total_time)
+labels <- plotdata %>% 
+    arrange(algorithm, k) %>%
+    group_by(algorithm) %>%
+    filter((k > lag(k)) | (workers == 8)) %>%
+    ungroup()  %>% 
+    mutate(
+        y = case_when(
+            (algorithm == "LocalLSH") & workers == 8 ~ total_time - 100,
+            (algorithm == "LocalLSH") & workers == 40 ~ total_time + 30,
+            (algorithm == "OneLevelLSH") & workers == 8 ~ total_time + 1500,
+            (algorithm == "TwoLevelLSH") & workers == 8 ~ total_time - 450,
+            (algorithm == "OneLevelLSH") & workers == 16 ~ total_time + 800,
+            (algorithm == "TwoLevelLSH") & workers == 16 ~ total_time - 350,
+            (algorithm == "OneLevelLSH") & workers == 40 ~ total_time - 130,
+            (algorithm == "TwoLevelLSH") & workers == 40 ~ total_time + 200,
+            # (algorithm == "LocalLSH") & workers == 40 ~ total_time + 30,
+            T ~ total_time
+        )
+    )
 
-ggplot(
+p <- ggplot(
     plotdata,
     aes(
         x = workers,
-        ymin = bestWorker,
-        ymax = best40,
-        color = algorithm
+        y = total_time,
+        color = algorithm,
+        shape = algorithm
     )
 ) +
-    geom_linerange(
-        position=position_dodge(2),
-        size = 1
+    geom_line() +
+    geom_point() +
+    geom_text(
+        aes(label=scales::number(total_time, suffix=" s")),
+        data=function(d) {filter(d, workers == 8)},
+        vjust=0.5,
+        hjust=1,
+        nudge_x=-1,
+        size=2.5,
+        show.legend=F,
+        color = "black"
+    ) +
+    geom_text(
+        aes(label=scales::number(total_time, suffix=" s")),
+        data=function(d) {filter(d, workers == 72, !((dataset == "Glove") & (algorithm=="OneLevelLSH")))},
+        vjust=0.5,
+        hjust=0,
+        nudge_x=1,
+        size=2.5,
+        show.legend=F,
+        color = "black"
     ) +
-    geom_point(
-        mapping = aes(y = bestWorker),
-        data=~filter(.x, bestWorker == best40),
-        shape=18,
-        position = position_dodge(2),
-        size = 3
+    geom_text(
+        aes(label=scales::number(total_time, suffix=" s")),
+        data=function(d) {filter(d, workers == 40, !((dataset == "Glove") & (algorithm=="OneLevelLSH")))},
+        vjust=0.5,
+        hjust=1,
+        nudge_y=-0.05,
+        nudge_x=-1,
+        size=2.5,
+        show.legend=F,
+        color = "black"
     ) +
-    facet_wrap(vars(dataset), scales = "free", ncol = 4) +
-    scale_shape_manual(values = c(2, 6)) +
+    geom_label(
+        aes(label=k, y=y), 
+        data = labels,
+        size=2, 
+        show.legend=F
+    ) +
+    # facet_wrap(vars(dataset), scales = "free", ncol = 2) +
     scale_color_algorithm() +
-    scale_x_continuous(breaks = c(1:5 * 8)) +
+    scale_x_continuous(
+        breaks = c(1:9 * 8),
+        expand=expansion(mult=c(0.15,0.15))
+    ) +
+    scale_y_log10() +
+    # scale_y_continuous(
+    #     breaks = c(272, 1412, 3351)
+    # ) +
     labs(
         x = "number of workers",
         y = "total time (s)"
     ) +
-    theme_paper()
+    theme_paper() +
+    theme(
+        panel.grid.minor = element_blank(),
+        panel.grid.major = element_blank(),
+        legend.position = c(0.8, 0.75),
+        legend.title = element_blank(),
+        panel.border = element_blank(),
+        axis.line = element_line(size = 0.3, colour = "black")
+    )
 
-ggsave("imgs/scalability.png", width = 8, height = 3)
+ggsave("imgs/scalability.png", width = 4, height = 2.2)

analysis/plot_system.R

@@ -2,62 +2,109 @@ source("tables.R")
 source("plots.R")
 
 plotdata <- table_sysmonitor() %>% 
-    filter(id %in% c(2719, 2720, 2721)) %>% 
+    ungroup() %>%
+    filter(hostname == "desktop2") %>%
     mutate(time = time / 60) # put time in minutes
 
-tot_net <- 1024 * 1024
+tot_net <- plotdata %>% summarise(max(net_tx)) %>% pull()
 
-ggplot(plotdata, aes(x=time)) +
+endtimes <- plotdata %>%
+    group_by(algorithm) %>%
+    summarise(end_local = max(time))
+
+plotdata %>% 
+  group_by(algorithm) %>% 
+  mutate(net_frac = net_tx / max(net_tx)) %>% 
+  filter(net_frac > 0.1) %>% 
+  summarise(start_comm = min(time), end_comm=max(time), comm_span = 60*(end_comm - start_comm))  %>%
+  inner_join(endtimes) %>%
+  mutate(span_local = 60*(end_local - end_comm), span_init = 60*start_comm)
+
+annotations <- tribble(
+    ~algorithm, ~from, ~to, ~text, ~y, ~ytext,
+    "TwoLevelLSH", 0.5, 1.4, "58 s", 0.25, 0.12,
+    "TwoLevelLSH", 1.5, 6.5, "307 s", 0.15, 0.25,
+    "TwoLevelLSH", 0, 0.45, "29 s", 0.75, 0.55,
+    "OneLevelLSH", 0.12, 1.2, "68 s", 0.25, 0.12,
+    "OneLevelLSH", 1.22, 6.1, " 293 s", 0.75, 0.55,
+    "LocalLSH", 4, 4, "Initialization: 2 s", NA, 0.75,
+    "LocalLSH", 4, 4, "Communication: 16 s", NA, 0.5,
+    "LocalLSH", 4, 4, "Local computation: 44 s", NA, 0.25,
+)
+
+percent_idx <- function(p) {
+    as.integer(quantile(1:n(), p))
+}
+
+ggplot(plotdata, aes(x=time, group=hostname)) +
     # CPU
-    geom_line(aes(y=cpu_user), color="darkorange") +
-    geom_line(aes(y=mem_used / mem_total), color="darkgreen") +
     geom_line(aes(y=net_tx / tot_net), color="blue") +
     geom_line(aes(y=net_rx / tot_net), color="steelblue1") +
+    geom_line(aes(y=cpu_user), color="darkorange") +
+    geom_ribbon(aes(
+            ymax=(mem_used / mem_total) + 0.02,
+            ymin=(mem_used / mem_total) - 0.02
+        ), 
+        fill="white"
+    ) +
+    geom_line(aes(y=mem_used / mem_total), color="darkgreen") +
     geom_label(
-        aes(y=cpu_user),
+        aes(y=cpu_user, hjust=0),
         label = "CPU",
         color = "darkorange",
         fill="white",
+        alpha=0.0,
         label.size=NA,
-        size=3,
-        data=function (d) { group_by(d, id) %>% sample_n(1) }
+        size=2.5,
+        data=function (d) { group_by(d, id) %>% slice(percent_idx(.99)) }
     ) +
     # Memory
     geom_label(
-        aes(y=mem_used / mem_total),
+        aes(y=mem_used / mem_total, hjust=0),
         label = "Memory",
         color = "darkgreen",
         fill="white",
+        alpha=0.0,
         label.size=NA,
-        size=3,
-        data=function (d) { group_by(d, id) %>% sample_n(1) }
+        size=2.5,
+        data=function (d) { group_by(d, id) %>% slice(percent_idx(0.99)) }
     ) +
     # Network
     geom_label(
-        aes(y=net_tx / tot_net),
-        label = "Net (tx)",
+        aes(y=net_tx / tot_net, hjust=0),
+        label = "Network",
         color = "blue",
         fill="white",
+        alpha=0.0,
         label.size=NA,
-        size=3,
-        data=function (d) { group_by(d, id) %>% sample_n(1) }
+        size=2.5,
+        data=function (d) { group_by(d, id) %>% slice(percent_idx(0.99)) }
     ) +
-    # Network
-    geom_label(
-        aes(y=net_rx / tot_net),
-        label = "Net (rx)",
-        color = "steelblue1",
-        fill="white",
-        label.size=NA,
-        size=3,
-        data=function (d) { group_by(d, id) %>% sample_n(1) }
+    geom_segment(
+        aes(y=y, yend=y, x=from, xend=to),
+        data=annotations,
+        inherit.aes=F,
+        linetype="dotted",
+        size=.4
+    ) +
+    geom_text(
+        aes(y=ytext, x=(from+to)/2, label=text),
+        data=annotations,
+        inherit.aes=F,
+        size=2.5
     ) +
-    facet_wrap(vars(id), ncol=1) +
+    facet_wrap(vars(algorithm), ncol=1) +
     scale_y_continuous(labels=scales::percent_format(), expand=expansion(mult=0.1)) +
+    scale_x_continuous(expand=expansion(add=c(0.3, 1))) +
     labs(
         x = "time (minutes)",
         y = "usage"
     ) +
-    theme_paper()
+    theme_paper() +
+    theme(
+        panel.border = element_rect(size=0.5),
+        panel.grid = element_blank()
+    )
 
+ggsave("imgs/system.png", width = 4, height = 3)